- Production can now wake a named human at three in the morning, by phone. Every one of our most serious production alarms fired as a warning into a chat room and then waited for somebody to happen to be looking, which is not an alarm, it is a note. There is now a real pager, built and shipped in four parts across the day. A critical alert sends a text message to the on-call roster; if nobody acknowledges within five minutes it places a phone call and reads out a spoken briefing of what is broken, written for the occasion, and it keeps calling in rounds up to six times until somebody answers. Acknowledging is a reply to the text or pressing any key on the call. A storm of alarms coalesces into one page rather than a hundred. It runs deliberately outside the systems it watches, on its own machine with its own credentials, because a pager living inside the failure it is paging about is not a pager; if its own home is what is down, paging degrades to the old chat room rather than to nothing. Seven of production's alarms are switched onto it to start, including the site being unreachable, turns failing, the database saturating and the daily model spend, and widening that list is a reviewed decision rather than a knob. The build itself refuses to go out if the pager is wired to nothing, or wired to an alarm whose name was mistyped, which are the two ways this fails silently.
- The launch wall board stopped being the slowest thing in production. The board that the room watches was rebuilding its entire picture from scratch on every poll, from every screen, and it had grown into roughly twenty database queries; with a few boards open it was about a sixth of all production web traffic at nearly a second a request, and it had pushed the whole site past its speed budget. Then a change to the numbers themselves quadrupled it to over three seconds. It was fixed in three passes. First, all the open boards started sharing one picture per period instead of each computing their own. Then the computing moved out of the website entirely, into a background worker that refreshes once a minute, so a viewer's request is now nothing but reading the answer already prepared. Finally the computation itself was rewritten: the same numbers now come from five statements on one connection instead of twenty-one over seventeen, taken at a single instant so every tile and both charts agree with each other, with nine new database indexes underneath. Measured end to end, 1300ms became 144ms on a mid-sized copy of the data and 1875ms became 507ms on a very large one, and the numbers were proved identical by running both versions side by side against one database with the clock held still.
- Access is now a level, not just a door: Read only or Full access, per area. Yesterday's viewer role could only ever answer "may this person act" for one single area, so there was no way to say "Talent Management, but read only" or "Alerts, but read only": a reviewer who should watch a list without editing it had to be handed the acting grant or nothing at all. Every grant now carries a level, and the level hides controls rather than pages, so a read-only holder walks the same screens and simply has none of the buttons that change something. Everyone who had access keeps exactly what they had. Around that, a careful set of smaller corrections, most of them about not lying to the person reading the screen: a panel no longer tells a full-access operator "your access is read only" during the second before it has actually checked, which is worst on the one panel that stops the assistant during an incident; a link into another area is drawn only when the person can actually open where it points, instead of promising a page that answers "access required"; the picker no longer offers a level on an area that has only one; and saving several changes at once now applies them in an order that cannot lock you out halfway through your own edit.
- The moment the job description is ready now stops the room, and Mira can press the button herself. The one irreversible step in the whole flow was being explained by a small coach mark that appeared under a button and quietly took itself away after twelve seconds, so the sentence explaining it was routinely spent on nobody. It is now a dialog in the middle of a dimmed screen, carrying exactly the same words, that has to be actively dismissed. Dismissing it does not fade it away where it stands, which says nothing about where the control it was describing went: it shrinks and flies into the button that stays behind, so the client's eye is handed to it. And a client who simply types "yes, go ahead" or "find me people" no longer has to go hunting for that control at all, because Mira can now press it for them. She does not re-implement approving and she does not decide readiness: she presses the client's own button, so everything that hangs off a real click still happens in exactly one place and cannot drift from it.
- How many questions Mira asks in one message is now her call. She was hard-capped at one question per turn, in four separate places. That is right for a relaxed client and wrong for a rushed one: somebody who types "what else do you need?" or "why so many questions?" still got the remaining gaps fed to them one per round. The cap is replaced by judgement. One focused question stays the normal case, but frustration, a hurry, an explicit "what is missing?", or a client answering several things at once is now permission to name everything still open in a single message. A multi-part ask goes out as plain text rather than as a card, because one card captures one answer and would silently drop the rest.
- The talent search stopped being an improvising agent and became a pipeline the agent steers. The search used to be a loop of an agent deciding its own next move, which meant the same brief could be searched differently twice and the cloud could rank differently from the bench. It is now a fixed pipeline, verified to reproduce the bench's own top three, with the agent back on top of it setting the search up rather than driving each step. Around that switch: quality scoring was retuned to the weights the team actually wanted and fit and quality now add up instead of one gating the other; the client's own reference images and what we know about their business feed the search again; a search that came back with nothing because the model returned an empty list of queries, which was five of seven live briefs, now falls back to a plain search rather than silently searching nothing; an agency whose quality grade arrives in an unrecognized form is no longer scored to zero; and a mistake that had every search sharing one cache key, so different clients could be served each other's results, is fixed.
- A client's answer mid-search now outranks the job description it contradicts. Proposals are scored against a copy of the brief frozen when the search started. When a client answers one of Mira's questions during the run, that answer reaches the talent and the live brief, but it never reached the scorer, so a client could release a requirement, a talent could offer exactly the scope that had just been approved, and the scorer would fail them on the very requirement that had been dropped. It happened in production this day: a client confirmed that placements were not guaranteed and that two items were out of the first month, and seven minutes later a native German speaker quoting exactly that scope at exactly the price was turned away for it, while two later proposals with identical commercial terms passed because they happened to word it differently. Answers given during the run now ride along with every proposal being scored, and where they disagree with the frozen brief the client's later word wins. They can only ever relax a requirement, never add one.
- A sentence a client typed into their own job description stopped disappearing. When a client hand-edited a block of the brief, the record still said the agent owned it, so the planner rewrote it about fifty seconds later and the client's typed sentence vanished from the page. An edited block is now theirs and the planner leaves it alone. The other half landed the same morning: a name or a role typed into the brief now also reaches the client's own profile card, which it never did, because the edit ran with that part of the system switched off and would have been discarded even if it had not.
- The database was rejecting roughly five hundred writes a day, and the error dashboard said zero. A single invalid character in a message, the kind that arrives from a paste or a broken encoding, made the database refuse the entire write. In production that killed about five hundred writes a day, seventy-four of which were whole client turns failing silently. Every one had been logged, and nobody could see it, for a second reason found the same day: expected everyday refusals, such as asking for a page that is not there, were being recorded at the same severity as genuine crashes, at a rate of about eighty-five per six hours, so real failures were buried under noise and the error stream read as healthy. Bad characters are now cleaned out before the write, and counted, so the rate is visible and points at whatever is producing them; expected client errors were moved down to a warning, and crashes stand alone again.
- Talent Management grew two lists it did not have, and the alert lists stopped being capped. Declines now shows, across every project, why talents are saying no, in their own words, and carefully separates a talent turning us down from our own decision to decline them, since both used to be recorded the same way and reading them together credited our choices to the talents. Opt-out requests is a durable queue for "a talent asked us to stop": that request used to exist only as a chat notification and a log line, so an afternoon nobody watched the channel lost it while the person stayed in the pool. Suppressing somebody stays a deliberate human act; what changed is that the request now survives a missed notification, and cannot be erased by a project being deleted. Separately, the Alerts worklist and the Big clients desk had both stopped at the newest hundred rows and told the operator to resolve some to reach the rest, which is asking a person to act on rows they have not read; both now page properly, so nothing is unreachable and no alert has to be dismissed to be seen.
- A talent's price now counts the moment they give it, not two questions later. Two optional questions sat between a talent finishing their answers and their proposal actually being recorded, so a talent who went quiet during those extras left nothing behind at all and their price and delivery date never reached the client. The proposal is now banked and reviewed the moment the last real answer lands, and the two extras merge onto it afterwards without being re-judged. Because it is banked earlier, the shortlist can be revealed while those extras are still arriving, so a personal note now routinely lands on a card the client is already looking at, and it re-composes that one card when it does. The talent still gets exactly one message per turn, so they read "recorded and cleared" followed by the next question rather than a goodbye followed by two more.
- The launch board learned to separate talent from clients, and to show frustration building. Some share of every day is talent who walk into the buyer product to pitch; they sign up, open a project and send messages, so until now they inflated every client number on the board and parked forever near the start of the funnel. They are now reported as their own pair of gauges and subtracted from every client number, with one careful clause: somebody flagged as a talent who then answered "I am here to hire" on the two-button card stays counted as a client, because that card exists precisely because the guess can be wrong. And the "Unsatisfied clients" tile, which could say how many but never whether it was building, gained a live sparkline and a 24-hour chart, counting distinct people rather than events so the chart and the tile can never print two different truths.
- Five things testers found in the client app, fixed. On a phone the shortlist could be swiped one way but never back, so the first talent card was unreachable by any input and its dot did nothing, a centring rule stranding a third of the deck off the left edge. The welcome suggestion pills, which are whole sentences, neither wrapped nor shrank on a small screen, so they bled off both edges and read as two lines of cut-off text. The popup that asks a guest for an email so a specialist can reach them had a close button, which was the one remaining way to dismiss it without leaving one. A talent card's button said "See full talent profile" on a card that already announces it is a talent. And clients on the simple search, who contact talents themselves, were never offered the job description to send along, so a line offering the download now sits directly above the cards.
- Two things were rolled back on purpose, and one had been quietly breaking research for weeks. A merge bringing nine job-description templates turned the build red on the front end, so it was reverted within the hour and will return once green, because a broken build breaks everyone's safety net. The second is more interesting: a fix from a previous day that stopped one guest's projects bleeding into each other had made the client's own identity write-only for guests, so the check that decides whether we can research a client's website read an empty field every time and deep research could never run for anyone not signed in. It was caught in preprod, where the assistant stored a client's web address and was told seventeen seconds later that none was on file, then told the same every turn for fifteen minutes. Reverted, with the original bleed reopened and the proper fix, a separate identity per project, written down as the next step.
A reliability day, 62 commits across the crew, and the theme is things that were quietly not working. Production got a real pager: a critical alarm now texts the on-call roster and, if nobody answers in 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, was moved off the request path and rewritten to be roughly nine times cheaper. The database had been rejecting about five hundred writes a day over a single stray character, seventy-four of them whole client turns, while the error dashboard read zero because everyday refusals were logged as crashes. Access became a level rather than a door, so a reviewer can hold "read only" on any area. The moment a job description is ready 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, which had cost one German speaker a real job seven minutes after the requirement was dropped.