The living reference for the Mira platform's agent system, plus the day-by-day build log. Pick a document below.
Fourteen agents behind Mira & Atlas — who each one is, what context it reads, what it produces, and how they are wired to one another. The turn lanes, the research chain, the concierge with STERLING & JUNO, PULSE the intent router, GAUGE the budget & timeline estimator, and the rules everyone obeys.
→Every agent's full system prompt, section by section, with a plain-English "why" for each part — the PM-readable companion to the architecture doc.
A Tuesday of 10 commits about a number meaning nothing without the thing standing next to it. A talent quoted forty dollars an hour and the closing note the client sent them said the job cost forty dollars. The reader who judges whether a quote is fair had never been handed the client's own record on the path almost every offer takes, and her written verdicts were being refused at the filing cabinet and dropped, ninety six in a fortnight. A shortlist promised in two hours turned up a minute after two hours. Signing back in after a tab had slept overnight said it could not verify the session it had just verified. And the strip of live matches stopped opening itself over the client's typing, which was the product answering the very question the experiment on that screen is asking.
→A Wednesday of 17 commits in which the price of a quote stopped being a fixed gate and became a judgement. The reader who weighs every talent's proposal now decides whether the price is fair herself, with the old bands written into her instructions, after being replayed on five hundred real offers and following her own rules on four hundred and ninety six. The checker that audits her now reads exactly what she read. A pause that holds a talent's first message until their morning now runs only on twelve hour hunts, and deleting an account no longer signs the person straight back in.
→A Thursday of 15 commits in which the client finally hears back. Messages we send to talents in the client's name now come back to an in-app inbox, with files both ways and Fiverr offers labelled as offers. A rule letting talents pick clients by nationality, race or gender is now refused by the system itself, after an instruction alone had failed. A live test run found six real bugs, from a delivery date a week early to a budget shown as half of what the client said, and all six are fixed. Two experiments were laid down switched off: a far cheaper model, and search reading the full brief.
latestA Monday of 21 commits, and the shape of the day is one way in not being a way in. A client's message now starts down two roads at once, because the queue that carries it quietly holds a ready job now and then while dispatching younger ones behind it. Three safety alarms could not be installed until the emergency they warn about had already happened once. A door had its key and its lock delivered to two different buildings. The third party every search and every talent message passes through had no window at all, and the probe watching it was writing its readings into nothing. Three separate changes went in and came straight back out the same day, all three queued rather than abandoned. The rest is craft: a booked call released on what the client settled rather than on half a sentence, a card that stops saying "off" about a product that is on, a deck that leads with its own recommendation, and five copy and phone fixes.
→A Tuesday of 38 commits, and the shape of the day is a confident answer from something that never looked. For forty three minutes the hire button returned an error to every client who pressed it, forty eight presses and no successes, while every test stayed green, because the test that drives it answers from a hand written stand-in that recognises a database query by spotting a word inside it. The report on that outage was corrected by lunchtime, its first numbers having been taken mid outage and printed as final, counting hunts and calling them clients. A key nobody issued opened a workspace; a search still running reported that it had found nothing; a count said eighteen over eight cards. Mira asked two of her own matches to send her their portfolios, because the notes she is handed listed her shortlist as bare names. A card painted its own idea of dark onto a room it cannot see. And a shipping gate passed an identity that was never filled in.
→A Wednesday of 19 commits, mostly about the gap between what a talent said and what Mira wrote down. A hundred and ten saved boundaries were read against the sentences behind them: a fifth were wrong, and most of those had already been approved. An empty save looked exactly like a deliberate clear. Taking back one rule deleted every rule. Ninety euros an hour was stored as dollars and then offered back to the talent as a cut. Around it, one operator opening the experiment board fired ninety five requests and three alarms, some turns were missing from every experiment without a single error, and visitors signed in to Fiverr are now signed in to Mira without being asked.
→A Thursday of 15 commits about doing a thing exactly once, in the order it was meant to happen. A client pressed stop twelve seconds after saying go and eight talents were messaged anyway, because the sender was working from a copy of a list instead of asking. A reserve of nine people waiting to be swapped in was removed, along with the timer whose only job was the swapping, and a hunt now approaches twelve at once. A deck of talents was being re-sorted twice on the way to the client, and crowned even when nothing had passed review. An hourly rate was recorded as a lump sum and then refused for being sixty nine times too big. And underneath it, a client's own record frozen once per project so every later judgement reads the same picture.
A Sunday of 6 commits, five of them on the same distinction: an empty row and a row that says nought are different statements, and a console that cannot tell them apart will eventually tell two different stories about one experiment. A stop now writes explicit noughts whether a person or the safety rule did it; a missing row reads as being AT the default rather than drifted from it; the sweep that puts an environment back learned to name rows for experiments that no longer exist, after the first version of that report shipped with the exact blind spot it was added to fix; and every kind of experiment whose baseline is a decision now has to state it out loud. The sixth is the specialist call card becoming something a client can ask for again, with the booked call kept in view, and the agent told to read the ABSENCE of a desk line as carefully as it reads the line.
→A Friday of 11 commits, all on one sentence: a baseline you worked out for yourself is a guess, and a console that prints a guess as a fact is worse than one that prints nothing. A dial reading "off, nought per cent" was sitting on a product that was fully on, and moving it to five per cent switched the behaviour off for ninety-five per cent of the work, so declaring the shipped baseline is mandatory now with nothing to fall back on. One button puts a whole environment back to what production runs, in one transaction, with a rehearsal first. A rollout guard that was reading the headline number learned to measure the distance from the baseline instead. And the mandatory reason came off the sliders after the first three real moves recorded "ttt", "ddddd" and "gggg".
→A Thursday of 66 commits, the largest day on this site, and nearly all of it is one idea: a switch is only real where somebody can see it and move it, and a version somebody was served is only real if it was written down. The experiment platform got its console and the console immediately revealed seven dials drawing sliders that reached nothing. The one control experiment whose whole job is to prove the measurement is honest had written zero rows in its entire life, across 139,883 recorded model calls, because writing them down was opt-in with four hand-wired ways to opt in. Mira inside Claude got its lock back and two lanes behind it. The employee console got its own service, its own pool and its own database user, two days after four admin page loads took 40% of one production machine's database capacity away from clients. A request that times out now actually stops working, rather than running to completion sixty-one seconds after the client gave up. And a hunt the client walked away from stops holding talents after forty-eight hours, while the talents stopped being told there is a clock at all.
→A Wednesday of 34 commits, and nearly all of it is the same idea from different angles: nothing can be measured, matched or trusted until it says who it is. Every database statement now names the service and the page that issued it, and anything slow gets written down, which is exactly what nobody could establish during yesterday's outage. A sick machine's measurements and its log lines finally share one identifier, so a single bad machine can be named instead of reading as a bad release. Every question on the waiting field carries the face of the talent who asked it. Where the work physically happens became a fact with a name, proven on five hundred real conversations before it was allowed to change anything, and locked so the search cannot let it go. And the connector inside Claude deleted its entire login apparatus in favour of one thing that says who the person is.
→A Tuesday of 27 commits, and the through-line is things that were quietly not happening. A maintenance job had been logging politely and doing nothing for days, because a letter was being compared to a picture of a letter. A production machine froze twice and served customers for an hour each time, because nothing had ever asked it whether it was alive. Two admin pages kept timing out a month after being paged, because paging bounds what you return and not what the database reads, which is now a rule with a 71-seconds-to-220-milliseconds measurement behind it. And the largest piece of the day is the opposite of all that: the waiting page became a field where a client watches real people move across six columns, shipped, sent to QA twice in one afternoon, and fixed both times before the day was out.
→A Monday of 25 commits, and the through-line is that an identifier is not the thing it identifies. Eight of them are one story: the card that books a call could not reach almost anybody, because a signed-in buyer is a number to us and the address their sign-in attests to had deliberately never been kept, so nearly every real client got a vendor pop-up instead of booking in the chat. A table of its own, then three sources read in order, then an inline question where none of them has one, then a version on the card so signing in can replace the frozen one on screen, then a move of the refresh to after the fetch it had been racing by 2.2 seconds, and finally a name on the calendar invitation, which until then read as seven digits and the host. A country filter held perfectly and a country is not a place, so a Rhode Island photographer was shortlisted for a ten-mile commute in Washington State. And 98.5% of the production database turned out to be the reports rather than the product.
A Sunday of 49 commits, and the through-line is that shipped is not the same as running. A closing line sent zero times out of two hundred and twenty thousand messages, four alarms that could never have fired, a kill switch that never reached the seam which actually drops people, and a whole event, reader, dispatcher and store machine deleted for having shipped and never once executed, replaced by the idea that a nudge is just a section of the instructions whose default wording is empty. The day's largest piece of work was going and reading what production actually holds: ninety two stored talent boundaries against the conversations they came from, thirteen of them wrong in four shapes nobody had written a rule for, and six thousand declines in which the person best placed to judge the match had already said it was wrong, in a column nothing has ever read.
→A Saturday of 4 commits, and the through-line is a row that tells the truth about itself. The readout could not say which version of the personality a project had been given, because the brief-building brain filed its record under the name of the brain that writes the reply, and the reply's own lane recorded nothing at all; both lanes now record their own sections and their own arm, and the arm is reported on the control too, since nobody was on it and it never ran are different facts. The experiment was bucketing a person while the report measures a project, so one client with five projects counted as five independent observations. Twenty five of twenty six named slots were missing from the page that exists to offer them. And the whole search journey became one clickable page for design review.
→A Friday of 12 commits, and the through-line is the name a thing answers to. The prompt registry named every brain after the file it lives in while the running system stamps a label, and the two had drifted on ten of fourteen entries, so the console's Mira tab showed a different agent and the sixty five thousand character prompt that writes every reply a client reads was not listed at all, which is why the voice experiment had been faithfully allocating people onto a prompt nobody was served. Mira gets a second personality at zero percent, on a mechanism that can REPLACE a section rather than only append to it, and a section will now show you its own text. A talent's price boundary was being applied to the two sellers standing next to them and not to the talent who set it. And a failed read of talent working hours had been holding every outreach until seven in the morning.
→A Thursday of 50 commits, and the through-line is that a switch is only real where it is wired. Forty seven prompts moved onto the dial that serves two versions of them, and then forty one of them turned out to be reaching nobody: correct output, believable numbers, nothing happening. Fourteen per-turn blocks that make up the largest half of what a client and a talent are told got an identity for the first time. Clients in nine countries now ride a lighter model lane decided from their own browser. The table that was sixty percent of the production database got a fourteen day retention and a metadata-only ledger that outlives it. A message waiting in a queue became a measurement rather than a subtraction done by hand. And shipping production became a named list enforced outside the repository, a day after a cancelled deploy migrated the database anyway.
→A Wednesday of 44 commits, and the through-line is one place to ask the question. Every quality question the product asks itself used to exist three times over, as the live prompt, as a transcription of it, and as editable rows, with nothing to say when they stopped agreeing; each is now one committed file that reproduces the real prompt byte for byte, with a test that fails in both directions. The Insights console was deleted whole and rebuilt as a registry of taggers on the reporting app, and the reader itself became an agent whose load-bearing rule is that an event it cannot decide is not a no.
→A Tuesday of 36 commits, and the through-line is a promise you can actually keep. Five talents were told their boundary was noted and nothing was stored, because every read and write keyed on a marketplace number that is nullable by design and missing on two threads in three; it keys on the username now, and where it still cannot key anything it fails loudly with no reply rather than sending a warm confirmation over an empty row. The rest of that flow got its four QA notes: a minimum that read back inverted, a second boundary that sounded like the first, an addition thrown away by a question about money, and a yes that saved nothing. Mira as an app inside Claude and ChatGPT went from skeleton to recognisable over eight passes, answering on her fast lane rather than the whole turn, wearing her own tokens, showing real talent photos, and no longer stacking the host's own instructions into the thread once a minute; she can also open more than one project now, which she could never do, and every new hiring need had been overwriting the last one. A booked call became a durable row instead of a single bit a takeover release erased. A project remembers the device it was created on. The brief stopped opening itself on a brief it had not earned, and then stopped shutting on a client who had just opened it by hand.
→A Monday of 44 commits, and the through-line is a search that explains itself, plus a Mira you can reach from somewhere other than our own browser tab. The live rail stopped blinking in three separate places, and the grading step that runs for two silent minutes now says how many it has reviewed of how many. What starts a refresh became a judgment the agent makes rather than a side effect of writing anything down, over a floor that refuses to search a project with no definition of its own, after a returning client's old research bought a hunt for fantasy-game talent on a logo brief. The shortlist re-ranks every round instead of freezing at the cap, which is why a client who asked for Spanish speakers had been reading a list where fourteen of eighteen spoke none. A brief that hires a person distils to the ROLE now, and the discipline check that should have caught it, dead and unreachable for three weeks, is back and metered. Talent standing preferences came back on with the hours somebody works, which weigh a little and never remove anybody. Mira shipped as a real app inside Claude and ChatGPT. The big-budget call returned as a scheduler card in the chat itself. And a proposal that cleared review and was declined a hundred and ten seconds later turned out to be a durable row losing an argument with a state column.
A Sunday of 28 commits, and the through-line is the phone finally being treated as a real surface. The document does not scroll on phones anymore: the shell is sized from the visible viewport and moved with it, which is what fixes the keyboard panning the page into dead space, the sent message jumping out of view, and rotation locking the layout on stale numbers. The brief drawer is gone and the chat home is two peer views under one toggle, with the browser's own Back leaving the document and both views keeping their draft and their scroll. The keyboard closes after every send, a docked card can no longer ghost-click the brief open underneath itself, and a message half-typed before signing in survives the redirect with its caret intact, on a landing that now boots once instead of twice. The question card waits for the first message and shows once per person, claimed when it is seen rather than when it is answered. The live search rail became one glance card that grows into the shelf, its talent cards carrying the grader's own reason on their face, and the finalist card learned to answer an ongoing hire in rate-at-scope, availability and repeat clients. A talent has working hours now, worth a couple of points and never a filter, and the twenty-minute chaser to talents is retired.
→A Friday of 9 commits, and the through-line is the link that knows where you were. Every ping we send about a project, in the inbox, on the phone and in the browser, was linking at the front door rather than at the project, so somebody tapping review shortlist landed on the lobby and, if they were signed out, on a blank guest home. All three go through one arrival hop now that reads the session and either lands them or carries the project through sign-in. The text needed a short spelling of that hop for a reason worth stating: past one segment the composer gives up the LINK rather than bill twice, so every text would have gone out linkless and nothing would have errored. The new public route was caught by the ledger that exists to notice one, within minutes of the push. Two chat-room pages about somebody confirming they sell rather than hire came out, while the worklist flag they announced stayed exactly where support already works it. And correcting one line of the brief now closes that whole SECTION to the agent, with the client the only one who can hand it back.
→A Thursday of 13 commits, and the through-line is the way back in and what got said while you were away. Signing in had three separate legs of one failure: the handler was awaiting a third-party account fetch inside its own ten second budget and burned thirteen production timeouts doing it, so it is a queued job now; the boot gate asked who you were with no time limit at all, so a stalled connection parked people on the boot screen forever; and the landing purged the guest cache it had just been handed, when the callback had already claimed that guest's project onto the account, so the cache IS the first paint and is re-keyed rather than dropped. A card on the production dashboard reported nought percent over twenty-eight thousand observations of a click that could not have existed yet, and every window beside it rendered shifted by the reader's own clock. A budget range arrives as two facts and the desk was paged on the first one and then frozen by the duplicate guard, so an open alert now amends itself. The new console gained the tester bug pad with automatic screenshots, a preference ruling can finally be taken back, and a chart of twenty-one daily cohorts stopped drawing seven of them and summing the rest into a grey blob.
→A Wednesday of 16 commits, and the through-line is measurement deciding what stays. Two search rules taught in the morning came back out by mid-afternoon with their numbers written into the reverts: the language a deliverable is in moved almost nothing across thirty-six replayed briefs even though the failure it reached for reproduces on demand, and a price rule bought a grader-internal correlation the client never saw while thinning the lists. The whole standing-preference feature went quiet behind one constant a day after shipping, deleted from nothing, because with the talent-facing half mute and enforcement still live somebody would be filtered out of every search with no way to see it or raise it. Hours earlier its last two pieces had landed: when somebody can start is now a date screened against the client's deadline rather than a rule sentence that could quietly grade them off everything. An account deletion failed nineteen times in an hour and was abandoned with the client's content still in the row. The console's main tab, which had stopped loading, is paged with its filtering moved to the server. And a four thousand dollar redesign was ruled small on its timeline and hard-routed to plain search, so clearing the money floor now sends a client to the specialist desk on its own.
→A Tuesday of 26 commits, and the through-line is who gets a say and who gets interrupted. A talent had exactly one all-or-nothing lever over what reached them, and can now state what they will and will not be brought, with three separate price floors because they are not convertible, and a human approves every one before it filters anybody. The way we asked for that boundary was rewritten the same day: a fixed line on every decline treated one pass as a request to be filtered forever, so Mira raises it only when somebody is telling us something should stop arriving, and stays silent on ordinary frustration. The strip beside the conversation went back to the real search at one round per refresh and became strictly append-only, so the third one is still the third one. Two thirds of the searches we were buying were thrown away unread while the counter watching for it sat at zero. And a brand-new seller stopped being deleted before anybody read their work, after a Spline brief lost the best match in the supply and the client was told nobody could do the job.
→A Monday of 40 commits, and the through-line is things that had shipped and were quietly doing nothing, each with a success-shaped signal sitting on top of it. The rail that searches while the client talks had its cheap engine gated behind a setting defaulted the old way, so the test group spent a day on the expensive engine at sixty-seven seconds and twenty-six cents a refresh. The button that hands that search to Mira was refusing almost every click and buying a fresh ninety-second hunt each time, which is why a client said it is not what I saw at the bottom of the search. And a grading call answered for all eighteen talents while the reader threw every row away over an at sign, under a log line that said done. With those fixed the cards carry their reasons again, the line above the shelf came back, and the order a search gives things up in is read from the client's own words instead of a constant. A logo brief that kept nine of a hundred and sixteen people and shipped a background remover as a finalist bought one core search at the finish, and the country penalty that shipped beside it was reverted the same afternoon. Allocation became a slider instead of a deploy, and an experiment arm with nobody in it stopped reporting that it had cost us nothing. And a hundred dollars an hour stopped reading as no money at all.
A Sunday of 37 commits, opening the week with two long branches landing at once, and the through-line is work already done being thrown away and asked again. A client narrowed the live search over eight rounds down to six guitarists in New York, pressed the button that hands it to Mira, and got a different list with a Nigerian seller on it. Five silent failures in that one run: the button started a fresh hunt from cold rather than continuing, the judge had no way to write that nothing had changed, finishing was read as a statement about what to search for so it deleted its own filters, the fallback then searched the marketplace for the literal first five words of the brief, and the stated New York requirement never reached the search at all. The rail now keeps the eighteen it showed you and fills the last places by giving one stated constraint back at a time, on a second engine with no model in it. A place the work physically happens in stopped being a preference to trade away, after a Swedish pharmacy errand shipped couriers in four other countries. Every judging turn in every hunt was erroring under a green test. And signing in was re-rolling which A/B group you were in, which at the live dial of zero meant every pinned tester dropped to control the moment they signed in.
→A short Saturday of 4 commits and a single arc: money that repeats is finally money. An ongoing hire commits a rate rather than a project total, and every surface that judges what a client is worth was still reading the total, so a six thousand a month retainer showed a dash in the console, counted zero in the reporting, and never reached the desk that exists for exactly that client, while a one thousand dollar one-off reached it easily.
→A short Friday of 6 commits at the end of a heavy week, and the through-line is what a client carries with them across a boundary. Somebody who spent an hour exploring as a guest and then signed in was announced, on their very next message, as a brand new first-timer: their hour of history only reached the account after the reply had already gone out, and the welcome flow outside the product keys on exactly that flag. It now moves across at the moment they sign in, in one statement that cannot leave the history in both places or in neither, so the first message after signing in claims nothing and the next real visit counts as a return. A brief printed to paper carries its own name and its page count on every inner sheet, its cover shrinks to fit rather than stranding one line on a near-blank page two, and a long list of references breaks where it should instead of jumping over whole, all four faults taken off five real briefs printed over four days. And the screen asking a client how they want to find somebody says what each door costs, a few hours or instant, before either card is read.
→A Thursday of 59 commits, and the through-line is judgement being taken away from constants that could not see what they were weighing, then one of the day's own answers being thrown out by its own measurements. A client wrote “must have ... knows english and spanish fluint”, and the single talent whose profile lists Spanish came back eleventh of eighteen, behind ten who do not, every one of them scored as meeting every stated demand: three hard-coded numbers weighed the misses and silence cost nothing. The grader now sets what a miss is worth itself, and is asked the question that decides the number, what breaks for this client if this talent is hired and the miss turns out to be real. The other half, telling the hunt what the client had committed to, was written four ways in five hours and every one bought compliance with something else; thirty-three briefs replayed five times said the build without it wins, so it came out at 23:21 with its numbers written down. The category a hunt searches is settled once per run instead of twenty-seven times, which had been failing nine searches in ten. Money got one reader instead of two, after an hourly rate was read as a fifteen dollar project and a yearly rate killed a live project outright, forever, while the client typed “comeon!!!”. Mira stops offering a one-click way past the questions that sharpen who comes back. And a client who deleted his account had his brief written back onto the row the scrub had emptied.
→A Wednesday of 46 commits, and the through-line is the search being wrong about things it had chosen itself. A robotics hunt reached a studio whose catalogue held the brief word for word at exactly the right price, and never looked at it: the step choosing which of a seller's services to grade picked by price, read their PHP and Android work instead, and dropped them. Every listing is now read, an hourly role is priced as what a month of it costs, or as the whole engagement when the brief says how long it runs, rather than as a bare hourly number that made every real hiring listing look fifty times over budget. The filter vocabulary now comes out of Fiverr's own catalogue instead of being invented, and the category we merely INFER moved off the main search, where it had been quietly gutting pools: one brief went from a hundred and sixty one people to seven. The week-long argument about the grader was settled in one hour by running four arms on one environment. Mira's account of the hunt became a real design in her own voice at half the length, the talent strip finally says it is a ranking, and the budget question is asked in the client's own unit, now including per year. Files stopped travelling through the web service. And four measurements that had been drawing a confident line while measuring nothing, one of them flat at exactly fifty for nine days, were found and fixed.
→A Tuesday of 37 commits, and the through-line is things that were quietly costing us while nothing reported a fault. A client asked outright what they wanted in a person, said “english speaking, based in europe”, and the search ranked six of its top seven against them: every condition about the PERSON was filed as a preference, and a preference could not move the rating at all, so a stated demand carried no weight whatsoever. It is now read off whether the client demanded it or merely floated it, in the conversation as well as the brief, and a failed demand ranks somebody below every compliant match, labelled, still shown. Grading, which is nineteen calls in twenty and four fifths of the bill, moved to a model a tenth the price, stopped waiting for its own slowest batch, and stopped being handed six searches at once in favour of one per round chosen after seeing what the last returned. Three doors that opened with somebody else's key are shut, and one that refused its own keyholder, an operator locked out of the console by the client cookie they were carrying, finally offers a way in. A fact a client deleted from their own profile stopped coming back on the next message, alongside eight more tester reports. And the live results rail now explains a pick before it sends anything, and asks first when the brief is not finished.
→A heavy Monday, 43 commits, and the theme is that the talent search stopped deciding who fits with rules. A dozen screens that deleted people before anybody read their work are gone: a Mexico brief lost all nineteen of its finalists to a country check comparing a two-letter code against the country's full name, another lost all fifty-one because it named a rare tool, and a blank field was being read as a mismatch rather than as unknown. Stated countries, languages and levels now steer where the search looks and reach the grader as evidence, and an empty shortlist is no longer an acceptable answer. Retrieval was also hunting for a bare common word and for a tool paired with itself. Finding people again cost real money and time, so the thinking spent per listing is pinned, a search is bounded by the clock, and no brief is paid for twice. The console can now ask “when” on every tab, says which talent the client actually pressed rather than which one we showed, and reads a client's conversation in English. And a run of reporting faults that had all been silently green: a warehouse frozen for eight days, production charts reading the development project, and a board plotting tokens per hour under a per-minute label.
A very large Sunday, 72 commits, and the through-line is that the product learned to hire somebody rather than only to buy a delivery. Every project until now was priced like a purchase: one total, one delivery date. A client wanting a designer two days a week for six months had that squeezed into fields that could not carry it, so the total became a made-up lump sum and the delivery date a deadline for work with no end. A work shape now runs the whole chain, read free on every turn by a step already running and settled once at brief approval: the client is asked in hours and days rather than employment terms, the rate is committed and priced in its own unit, the brief closes on a start date, the search rewards repeat clients and capacity over one-off volume, and the talent hears the real shape and quotes per period against it. Beside it the fee and the money that merely passes through a talent stopped being summed, which had priced ambiguous jobs wrong in both directions. The search got a sharp correction: a client saying "$500" was being read as refusing anything cheaper, deleting the affordable talent the search existed to find, and the reviewer that blamed the price band for every empty pool now tells a real price problem from a marketplace that returned no prices at all. Admin stopped counting the shortlist going up as the client picking somebody. And the test suite, the slowest stage in the pipeline, was spending five hundred of its eight hundred seconds waiting on a machine it could never reach, with nothing ever going red to say so.
→A short Friday, 4 commits, all of them closing the loop on the week's two big pushes. The talent grader turned out to have been promised a full profile in its instructions and handed almost nothing in practice: sixty of ninety-seven packages carried only boilerplate, so for two thirds of a pool it was judging capability from a line that said nothing, and it correctly refused to award anybody top marks. It now gets the portfolio, the skills, the real description and, most sharply, language proficiency rather than bare codes, which matters because the brief that started all of this asked for native German speakers. A second search fix caught the round meant to widen the pool silently tightening it, flipping "speaks German or English" into "speaks German and English" through the one field the safeguard could not see. And yesterday's atomic reveal claim got its other half: the losing job stopped doing everything that comes after the reveal.
→A Thursday, 43 commits, and almost all of it is one sustained investigation into why the talent search kept coming back empty. The answers were not subtle. We were handing the meaning-based engine a sixty-character keyword instead of the brief, and it was inventing an entire fictional project out of "google drive" and matching on that. The specialist engine built for us three weeks earlier had never once been called, and the leg named for briefs was plain keyword matching. We asked for talent speaking a language code we made up, which deleted three hundred and twelve of three hundred and fifty-two candidates. We deleted anybody who had left their language field blank, all two hundred and seventy-two of them on one round. We deleted an American developer who could build exactly the app described, for not speaking German. And the reviewing agent could turn a real shortlist into a blank page by opinion rather than arithmetic. All of that is fixed, the two halves of the search now hold one conversation instead of two independent guesses, grading costs a fraction of what it did, and an empty hunt finally says why, which matters because one had pooled zero out of two hundred and fifty-seven people during a twelve-minute upstream outage where every call returned a healthy 200 OK. Elsewhere: a text message and a browser notification for the client who closed the laptop, fifteen seconds of dead air before the talent cards appeared, and a reveal that had fired up to three times for one client.
→A Wednesday, 31 commits, mostly about knowing things. Nearly nine in ten clients were recorded as arriving from nowhere in particular, which is not behaviour, it is a note we threw away at the door: about twenty buttons across Fiverr's own pages tell us where somebody came from, and the branch handling all but three of them discarded the label before anything recorded it. The team can now read our own conversations back through the assistant itself, and cross two layers that have never crossed, with every cell carrying the number it was counted out of and every exclusion stated. The tool that reads our numbers back to us got a named per-person lock on its door and stopped riding the product's release. A new board answers what none of the others could, how much room is left before the next step in traffic hurts, with every ceiling line drawn from the setting that provisions it so it cannot drift. Production can be put back now, code rolling back while the database stays put. Every deleted guest had been counted as a signed-in client since the fourth of August. A talent search that found nobody stopped giving up in silence, and went from zero finalists to eighteen on both of the staging briefs that had died that way. And the shortlist became one scrolling deck at every width, with one press to reveal everybody.
→A very big Tuesday, 52 commits, and the headline is that we stopped renting the tool that reads our own numbers back to us. Eleven dashboards and a hundred and sixteen charts now run on our own reader, on its own address, fetching the dashboard definitions from the analytics project while it runs so a chart edit no longer needs the product deployed. The half worth reading is what building it found: a headline number labelled current was showing a value from twenty-four days earlier because nothing had asked the database to sort, a chart was silently dropping a third of its rows on every load, twenty charts died outright when a reader touched the date control because a funnel stage name was being read as a timestamp, and every daily label was drawn in the reader's own timezone. Elsewhere: a large client handoff wrote nothing for twenty-six hours while eleven clients were told a specialist would call. The talent search stopped deleting the specialists the brief was written for, in three separate places, and every filter we state is now enforced by us rather than hoped for. A talent setting their own terms shipped, grew four fixes and came back out the same day. Signing in became a gate rather than an offer, and two browser tabs stopped fighting over one project.
→A Monday, 16 commits, mostly about the document the client walks away with. Printing a job description stopped needing a browser reserved, held and handed back, and became one request in and one file out, after production spent two days refusing seventy-two of them while its machines reported themselves half idle. It cost eight clients a document and withheld three promised messages, and the fix took thirteen review rounds, three of which found defects introduced by earlier rounds. Underneath that, six of the nine designs could not be saved at all, because the database still held a hand-written list of the original three and refused everything newer, killing the whole print job rather than just the preference. The big client flag stopped freezing the client it was about: it notifies now, and nothing stops. The plain search learned to show more than three people, twelve of them, three at a time, with a comparison table researched rather than guessed. And in the background: the release board stopped crediting a new version with the old one's day, and nineteen abandoned services were found billing for pull requests closed weeks ago.
A long Sunday, 34 commits, and the through line is that heavy work stopped happening while somebody waits for it. Printing a job description takes a real browser about nine seconds, and it was being done inside the click against a ten second ceiling, which on staging meant one talent pressing one button five times before it worked. Every remaining place that did it moved off the clock today, for the talent, the operator and the client alike, along with the long tail of races that comes of telling a browser later that its file is ready. In the same area the talent and the client turned out to be reading two differently designed copies of the same document, which now cannot drift apart, and picking somebody off the plain search results now actually sends them the job rather than opening an empty inbox. The approve popup stopped announcing itself uninvited, and Mira lost the ability to press it on a client's behalf. The concierge opens with nine talent instead of six and can reach eighteen instead of twelve. A new board reads six measures either side of a release. And underneath all of it: one unprintable character had been breaking entire database reads in production for real clients, file downloads stopped being carried through our own service, and three alarms that had been paging humans about nothing were quietened, one of them holding the same incident open for twenty five days.
→A quiet Saturday, 4 commits, all of it spent turning the Insights area from something that describes our conversations into something that measures them. Every project now carries plain observable outcomes beside its labels, facts read off what actually happened rather than anything a model was asked to judge: the client left without ever searching, the search found nobody worth shortlisting, the talent never replied. Each label is then scored against those outcomes statistically, with a confidence interval and a letter grade that keeps the size of the effect, the certainty and the weight of evidence as three separate things, because a model asked what was to blame will always find you an answer. A new probe replays each conversation as a growing prefix to find how early a problem was visible and how steady that signal is. The set of questions gained version control, since changing the instrument halfway through invalidates the measurement. And the assistant can now edit and retire checks rather than only propose them, apply a batch in one press, and keep one continuous conversation across reports.
→A shorter Friday, 16 commits, on capacity. The production database had been filling at nearly three gigabytes a day, and eight and a half of the nine and a half gigabytes in use turned out to be one thing: the full text of every exchange with a model. That heavy part moved out to cheap bulk storage with the database keeping a light pointer, everything already written is being carried across in the background, and both sides now expire after ninety days, with the disk itself finally charted and alarmed. Background jobs stopped being sorted into the fast lane by fallthrough, which had quietly enrolled six job types in a thirty-second promise that all six break. An alarm that fired fifteen times in four days turned out to be counting buyers closing browser tabs. And the Insights assistant learned to see every agent at once, point at the exact rows it means, and survive an answer longer than its own size limit. The day closed on a deploy that finished green having shipped nothing, and could not be put right by pushing.
→38 commits, in two clear halves. One is a brand new Insights area that reads our own conversations back to us: run a report, and every conversation is checked against a list of questions a teammate writes from the console, with each answer stored beside the quote that proves it. Over the day it grew a proper answer shape so results can be counted, a second surface for the talent-side agent read per thread, then every agent through its own recorded working, then charts, then a history across past reports, and finally a floating assistant that answers questions about the run on screen, points at the exact rows it means, and offers to add the question it thinks is missing. The other half is production telling the truth: the channel all sixty-eight production alarms routed through had never delivered a single one since it was created nine days earlier, including four days of the admin console being completely down; previews stopped setting off the shared environment's alarms; talent replies became measurable end to end; and the funnel board went from 8.5 seconds on a page request to 265 milliseconds off it.
→A reliability day, 62 commits across the crew, spent on things that were quietly not working. Production got a real pager: a critical alarm now texts the on-call roster and, if nobody answers within five minutes, phones them and reads out a spoken briefing of what is broken, from a machine deliberately outside the systems it watches. The launch wall board, which had become about a sixth of all production web traffic and pushed the site past its speed budget, moved off the request path and was rewritten to be roughly nine times cheaper. The database had been rejecting about five hundred writes a day over a single stray character while the error dashboard read zero. Access became a level rather than a door, so a reviewer can hold read-only on any area. The brief-ready moment now stops the room in a dialog that flies into the button it explains, and Mira will press that button herself when a client says go ahead. And a client who answers a question mid-search no longer has their own answer used against the talent who followed it.
→The day after launch, and it reads like one: 51 commits, most of them closing something that had been standing open. A talent who wanders into the client chat is now asked which side of the marketplace they are on rather than guessed at, and their own answer, not ours, closes that one conversation while leaving their account untouched. A security sweep found that the route which spends model money on the public site had a lock whose off-state was "allow" and sixteen siblings with no lock at all, and left behind a test that fails the build the next time someone forgets. The console gained the operator tools it had been asking for: a "Brief finished" rung and exit percentages on the funnel, a filter on every column whose values repeat, a blocked-users tile, a permissions table that shows a person rather than a list of grants, and a viewer-only role that reads a conversation and touches nothing. On the talent side, one canonical conversation per person per search ended an alarm firing fourteen times a day, and Mira learned to answer "ok" with one line and "what are you" with the truth.
→Launch day. 64 commits across the crew, and the testing window closed: real clients, real talent, real messages, with the production database given a delete lock the cloud console itself honours and a daily copy kept outside it. The day's own incident came within the hour: one in ten job descriptions failed to render into a PDF, and because the attachment was best-effort the opening message went out anyway, telling the talent a document was attached when none was. It now renders once per send rather than once per person, retries, and withholds the message rather than sending one that lies. Around it, a new Funnel board draws every project's journey and where it stops, in projects and again in dollars, and the TV dashboard learned to open the list behind any number. Mira now spots a talent who wandered into the client chat, a hand edit to the brief wakes the planner too, and the scout judges fit on the work someone actually sells rather than their headline.
A very full Sunday, 41 commits across the crew, on being able to watch the thing run and on messages that actually arrive. A launch-day wall board went up in the admin console, realtime to the minute and openable on a TV with a one-time code, and the dashboard that explains why an agent did what it did finally reaches production. Underneath it, a serious find: the reminder and the four closing messages to talent were written and saved but never handed to Fiverr, silent in production while looking healthy in test. Mira dropped her sign-off, a rejected offer stopped being answered with a receipt, Mira's Choice stopped crowning the best-decorated seller before anyone read the fit, and the scout now carries stated preferences and the country and language screen onto every engine. The exported document lost its invented client logo, nobody is handed a brief they did not ask for, and every pill and chip in the app finally sits on its optical centre.
→A short Friday, 12 commits, spent almost entirely on the document a client walks away with. Four different places handed out a PDF of the job description and they did not agree, one of them an old template with a different design, so all four now produce the same document with the Mira mark on every page. The page check stopped predicting where the printer breaks and started reading the finished file, so 7 of 8 multi-page briefs now end on a full page instead of 3. A site that blocks robots is understood as a real business, a client whose name we already had is no longer welcomed back on their first ever message, and the "search Fiverr instead" link finally types what the client asked for.
→The biggest day yet, 89 commits across the crew, mostly about handing control to a human: the new console became the place the team works from, gained a big-clients desk, and can now pause or stop the assistant and answer a client directly from the project page. A flag goes up from $1,000 and at $6,000 the system refuses to search and hands the client over by itself. A real buyer who said "ILS 3,000" had been pushed to the cheap lane. Every search is now written down with the scout's own reasoning, the brief writer stopped being told the wrong industry for six of eleven categories, the reporting warehouse was pinned to the version each environment really runs, and the app took on the Fiverr brand faces.
→The biggest day in weeks, 54 commits across the crew: a new admin console opens with an Explorer that puts a person, their projects and a whole conversation one click apart, a proper worklist for raised flags, and a talent management dashboard built on the analytics warehouse that refuses to show unknown as zero. The scout stops ranking a must-missing person above a compliant one and starts showing the package tier that actually does your job, research stops turning a dead parked website into a client's market, and the client app takes another reviewed-wording pass.
→A focused Tuesday, 9 commits, mostly making the talent search honest about money: cards now show the package that truly fits your budget instead of a seller's cheapest floor, there are no more $0 cards and no bargain-priced sellers passed off as an exact match for a far bigger job, and the scout stops throwing out good people over a word mismatch. Alongside it, the client app got its full reviewed wording pass.
→A short Monday opening a new week: 3 commits on copy and clarity. The house writing style now bans the long dash everywhere a client or talent reads, held in place by one back-end cleaner and a front-end build check plus a full sweep of the existing wording; the talent hears exactly one warm message when their proposal lands; and the guest sign-in popup reads in three clear levels with its opt-out as a plain link.
The biggest day since the production push: high-value client takeovers now feed a customer-success handoff into the analytics warehouse, the admin side gains real moderation power — block, unblock, delete and a full audit of every talent conversation — sessions fail softly instead of silently, budgets lock to whole dollars, and the crew brings up an external one-call entry into Mira, a curated VIP shortlist for the scout, and a full data-and-analytics platform for the pre-launch environment.
→A heavy craft-and-quality day: the client app becomes fully translatable — some 830 strings across fifteen areas — matching turns fairer with coarse quality tiers, a binary pass/fail on offers and standing-aware "Mira's Choice", the brief becomes something you truly edit, Mira's questions and widgets get clearer, sessions stop dying silently, and the crew filters inbox automation, lands an analytics events plan, and makes the site deploy atomically.
→A big scout-and-quality day: the talent scout becomes a disciplined headhunter — core-profession fit gating, discipline-anchored tiers, must-have screening and hard location and time-zone gates — searches through a third engine aimed at a focused shortlist, the match cards show the package that actually fits your budget and deadline, STERLING lands a batch of polish, big clients are told during a human takeover, first briefs stop inheriting prior context, and the crew adds error boundaries and splits oversized live messages.
→A quiet housekeeping day after the scout push: the test guarding the new best-fitting-package cards moves to sit beside the code it covers and stays green under strict type checking. No user-facing change.
→A tidy correctness day: money now reads the same everywhere — Mira picks up the exact currency you name, tells you the brief will convert it to dollars, and the brief prints a clean range with no stray "USD"; the results-page sign-in button just signs you in; the compare table gains honest "Mira's take" and "Still open" rows and drops the flaky Responds-in band; and switching accounts leaves nothing behind.
→A heavy craft-and-quality Saturday: the chat box is ready the moment you land, editing the brief's lists feels like a real editor, Mira's docked question reads cleanly in bold, talent cards read like a person rather than a listing, a genuine accessibility pass lands screen-reader announcements and keyboard-friendly tables, sessions heal after a timeout, and a raft of reliability cleanups run under the hood.
→A calm Sunday closing out the week: the brand-look intro learns when it has really been seen, the results lock note is restored to its dictated wording, Saturday's accessibility-and-safety cleanup gets its follow-ups, correcting your website now resets "About You" as well as the brief, and each talent card is priced from the exact gig it shows. The week ends settled.
The biggest day of the week: talent messages switch to real-time delivery, the results cards get a full redesign with delivery, price and one clear Mira's Choice, outreach speaks and signs as Mira with the brief attached, a scoped talent-platform button appears once matching starts, brief exports adopt the site design, and deep monitoring lands under the hood.
→The heaviest day: the talent conversation is rebuilt into three calm blocks with a real review before any reply, PORTIA joins as the fifteenth agent to write each match's human pitch, search leads with proven capability, talent's own files are read and scored, and the crew lands a full match-card redesign.
→The biggest day of the week: the talent scout turns into a relentless, source-aware headhunter — one door per source, a floor of twenty-plus, and deep reading of each person before it ranks them — budgets can now be a range with a soft floor, PORTIA's pitch gets tighter and reaches guests, briefs export pixel-identical to the app, and the crew lands a card redesign, a broad refresh-and-restore hardening pass, an admin scout-replay tool, and the first foundations of a data & analytics platform.
→The honesty day: the scout stops trimming and reads every candidate in full, cheap talent stops being punished for its price, proposal reviews return a plain pass or fail, foreign budgets convert at a live rate, the cards drop their invented stats for real standing, Mira says she's an AI in her first line, the brief exports as a finished PDF, and the crew brings up the first production environment.
→A steadying day: the match-cards page stops needing a manual refresh to appear after a reveal and self-heals a stale view, the downloadable brief renders the real designed document again in preview instead of an old template, and the logo tile settles into a proper card once the logo paints.
→The big feature day: high-value clients gain a human takeover — the team can pause the assistant, step into the Fiverr conversation by hand and hand it back later, with the client kept calmly informed — while the offer reviewer starts reading the full talent thread, the scout exposes its complete evidence and rewards proven must-haves, launch dates stay honest, and the brief, chat and talent cards each get polish.
The recognition day: the assistant greets you by name, the little brief tab shows live progress and opens to its checklist, the talent cards get a redesign, a final skippable search question shapes who you see, and the live-update plumbing gets calmer and sturdier.
→The two-way day: talk with talent flows in both directions with your files attached, the buyer app fits your phone, the assistant greets you plainly and researches the site you give it, split budgets add up correctly, and the question cards get friendlier.
→The brief-true day: talent search is driven by the brief you can see, asking for advice stays on-product, your match cards survive a sign-out, research stops touching competitors, the number card retires, and the brief links back to your business.
→The Mira day: the app takes its assistant's name, your session lasts a month and fails softly to guest mode, the assistant learns your brief edits and stops repeating itself, stays honest about impossible deadlines, reads your files in full, hires people as people, and talks with talent through the real inbox.
→The craft & routing day: swipe through matched talent on your phone, share any candidate by link with working back & forward, see why each one fits, browse a tidier files gallery, know that deleting your account truly clears you out — plus a set of small comfort fixes.
→A quiet Saturday safeguard: locking a brief now fully freezes it — the remove control on each file and reference tile is switched off, so nothing on a read-only brief can be deleted by accident.
→The go-live day: the pre-launch preview site steps onto Fiverr's real sign-in with its old blockers cleared, outreach to the same talent for two different clients stays cleanly apart and speaks in your voice, new talent replies surface the instant they arrive, and the finalist card gets a tidy-up.
A disciplined day: the assistant learns your business before your problem, the brief locks down the few facts it can't do without (name, budget, timeline), budgets count only when you agree, filler "Got it —" openers are gone, and sharing & progress-bar fixes land underneath.
→The safety day: uploads are virus-scanned before they send and files drag straight onto the chat, GAUGE joins as the fourteenth agent to ground budget & timeline, the concierge reaches wider, the brief keeps getting cleaner — all on a stronger model with steadier plumbing.
→The transparency day: you can watch MIRA think as its reasoning streams live into a calmer, Claude-style chat, only the files that actually matter reach your brief, the client's own homepage gets pinned into it, and every attachment becomes a real clickable link.
→The plumbing & discovery day: SCOUT learns to find talent on its own through Fiverr's official search door (the old scraper retired), you curate exactly which research sources land in your brief, its final touches always arrive and build faster, and the behind-the-scenes machinery gets sturdier under load.
→The cleanup & correctness day: deleting a project finally clears it away everywhere, the pictures in your brief stay put, deep research waits for the client's real website before it digs in, your dollar figures stay in dollars, and stale reference tags tidy themselves up.
→The personalized day: the assistant greets you already knowing your business, your brief leaves as a properly designed document, talent's questions arrive in calm batches, budgets are pinned to live market prices, and a run of correctness fixes make the offer cards and briefs read cleanly.
Edit the brief in place, MIRA matches how technical you are, a smoother refresh-proof search, and the browser moves into its own sealed, isolated service.
→PULSE joins as the thirteenth agent — sensing window-shoppers and routing them to plain search — JUNO judges offers against your real brief, declines now come with a reason, and your uploads can finally be read.
→The big one: your uploaded files can finally be read — parsed, summarized and handed to the assistants. Testers get a screenshot-backed Debug log, closed projects keep their thread, and the assistants ask less while knowing more.
→A polish day: the brief now wears the client's real logo and sheds its empty sections, the message box calms down, .json & .md uploads are welcome, the concierge says exactly what an offer is missing, and your guest work follows you in at sign-in.
The brief and the "About You" page share one polished header with Share built in, the assistants get sharper about budgets, shared links, tools and one-off-vs-ongoing work, files and generated images finally stick, and a long list of rough edges gets smoothed.
Watch the brief compose itself live, STYLO reads the client's real brand colours, HERALD narrates the match, and your guest work follows you into your account.
→Real Fiverr sign-in, inline colour-coded change confirmations in chat, refresh-safe concierge, and research that remembers the company.
→A craft day — a responsive chat composer, the brief dock that opens itself, the new "About You" look, and admin Google sign-in.
→Broken images fixed, approved offer terms reach the client, budget-as-a-range, and the database now updates on every deploy.
→The brief now writes itself live, a calm one-screen waiting room, JUNO joins as the offer scorekeeper (twelve agents now), STERLING renegotiates low offers, and no brief starts without a budget & timeline.
IRIS becomes the identity planner lane; one tagged notes channel to Mira; STERLING lands offers the regex can't parse.
→Concierge auto-reveal, STERLING's run-wide answer memory + MASON/IRIS ingestion, the fiverr_handoff stage — and the Atelier redesign begins.
→The Mira rebrand (talent/client), research-chain v2 with owner-placed imagery, un-gated planner lanes — and this docs hub goes live.
→The admin control room — PM/CS dashboard, operator takeover, auto-flag + ops alerts — and suggested future projects replace the project map.
→Timed proposal collection in shrinking windows, talent waves & auto-replacement, a search fallback — plus admin debug scaled.
Observability foundation, the "About You" identity file with branded PDF export, and the deep-research imagery pipeline.
→Agent architecture overhaul — MASON becomes the sole brief author and the crew gets its names.
→Ocean11's last feature day (brief-as-dock, agent-decided checklist) — and the big port into recruiter begins.
→The port lands: integration tests on a real DB, infra reconciliation from a live terraform plan, the buyer SPA pages.
→Live Fiverr discovery (SCOUT) narrated over Pusher, the 10 KB-cap resilience work, and HUE's PDF design job.
→A quiet plumbing day: the brief board's edits and drag-and-drop rewired onto clean data hooks, and save failures now surface instead of hiding.
→Craftsmanship: shared types consolidated into one package, the brief's drag logic extracted into a reusable hook, and a compiler guard against unhandled cases.
→The headline: a live seller platform with STERLING the conversational negotiator, plus full cost & I/O telemetry.