Operations
It worked in month one.
Month four is where an agent goes wrong quietly. A vendor ships a new default into the config overnight. A webhook starts returning a five hundred and the dashboard stays green. A retry books the slot twice. Nobody is watching the destination, and the first person who notices is a customer.
The Production Operations Retainer is the watch. From $3,500 a month, for a system we built or audited: watchdogs on the seams that fail quietly, a receipt on every consequential action, a nightly reconcile that pages a person, a monthly report with numbers you can audit, and one live fix window a month. Month to month. The report is the receipt.
Why month four is different
The vendor rewrites your config overnight.
A platform ships a new field with an active default nobody chose: recording on, retention unlimited, a filler phrase that stacks on top of yours. Nothing errored. The agent simply behaves differently on Tuesday than it did on Monday, and no one on your side changed a thing.
The destination goes quiet. The dashboard stays green.
The post-call webhook fires, the write fails, the handler returns two hundred, and the transcript is gone with no redelivery. The code path throws no errors while the table sits empty for weeks. The most expensive failures are the silent ones, and a dashboard cannot see them because it watches the process, not the row.
The model gets swapped, and the agent has to still be the agent.
Providers retire models. Prices move. A better one ships. The system has to change bodies without losing what happened, what it committed to, and what it is allowed to do. That is the category we build in: the history, the commitments, and the receipts live outside the model, and the authority to act is enforced in code.
The watch, on our own front desk
Our front desk is a production voice agent on a public number. This is its watch log for the first week of September, one line per mechanism, each with a receipt. A system under the retainer gets the same log with the same receipts. How this log came to exist:The Far End of the Pipe.
- every six hoursThe live agent configuration is pulled from the vendor and diffed against the committed baseline. A field changed by anyone but us pages a person; a deliberate change is accepted only after the pinned behaviors pass again.
- after every push8 pinned behaviors re-run against the live agent: pricing quoted firm and never discounted, caller instructions treated as content and never as commands, a recording objection honored without argument, the email read back before it is captured, the lead captured before the booking call, and three more.
- 2026-09-03The public door moved off a tunnel that had closed silently twice. 8 tool endpoints were rewritten and read back from the vendor before the change counted.
- 2026-09-04The model behind the agent changed after a bake-off: 24 of twenty-four pinned-behavior evaluations green with zero retries, and faster than the incumbent on every latency figure.
- nightlyTranscripts are counted against calls and receipts against actions. A missing one pages.
What the watch does
- 01
Watchdogs on the seams
The public door is probed from outside your network, not inside it, and a person is paged on the first failed probe. The agent configuration is fingerprinted, so a vendor-shipped change is caught the night it lands. The webhook registry and the destination tables are checked for recency, because a table that stopped growing is the earliest signal you will ever get.
- 02
A receipt on every action
Every send, booking, write, and spend lands a row before the provider is called, so a retry can never become a duplicate and an uncertain send becomes an in-doubt row a person reviews instead of a blind resend. The receipt is what makes the monthly report a report instead of a story.
- 03
The nightly reconcile
Transcripts are counted against calls, rows against events, receipts against actions. A missing one pages a person that night. A dashboard would have shown green; the reconcile shows the gap.
- 04
The monthly report
What the system did, what it cost, what changed, and what we changed. Reporting from the systems we run, billing reconciled against the receipts, and model routing that cuts inference spend without cutting output, each as a number you can check.
- 05
One live fix window a month
Scoped changes ship under the same gates our own builds pass: a failing test first, adversarial review, a live verification after the deploy. The diff is in the report. Anything larger is a build, priced as one.
What it is not
- Not a help desk for a system we have never seen. The watch starts from a baseline, and a baseline comes from a build or an audit.
- Not a promise of zero incidents. Security and reliability are a watch we keep, never a state we claim; the promise is that a person hears about it first, and you see it in the report.
- Not a rewrite lane. A change bigger than the fix window is a build, with its own price and its own gates.
We run it on our own systems first
Our front desk is a production voice agent on a public phone number. Its door is probed from outside our network, and a person is paged on the first failed probe and again every six hours while it stays closed. Every landed call leaves a receipt; a nightly reconcile counts transcripts against calls and pages when one is missing. The agent was audited against the same43-probe catalog we sell, and the first run found 3 severity-one findings on our own line, fixed the same night and verified by re-running.
Underneath it, the cognitive runtime that carries the fleet takes a core state snapshot every three hours and a full one nightly. The latest core snapshot holds10 tables, every file hashed and matched against its manifest on the day this page was published, and a cold restore into a fresh database was rehearsed on 2026-09-04, table by table. That is what the retainer's watch looks like when it is pointed at ourselves: the seams named, the receipts kept, the restore rehearsed rather than assumed.
Questions buyers ask
Which systems qualify for the retainer?
Systems we built, and systems we audited. The watch needs the seams named and probed once, and an audit is how that happens on a system we did not build. We do not take a retainer on a system we have never seen; the first month would be an audit anyway, so we sell it as one.
What is in the monthly report?
Every consequential action the system took and the receipt it landed; every watchdog transition and what was done about it; every miss the nightly reconcile found; every configuration change the vendor shipped into your agent; the model routing changes and what they saved; and the diff from the fix window. Numbers first, then what changed, then what it cost.
What happens when something breaks at three in the morning?
The watchdog pages a person on the first failed probe, and again every six hours while the seam stays closed. On the systems we run that person is us. On yours the pager goes to whoever the addendum names, and we are on it. Incident work outside the monthly fix window is scoped on the spot at a fixed price, and you decide.
Does the retainer include model and vendor costs?
No. Inference, telephony, and hosting are billed to you directly; you own the accounts and the keys. We report the spend line by line and route models to cut it without cutting output, and the report shows the difference.
Can we cancel?
Any month. The last report is the handover: the watchdog configuration, the reconcile queries, and the receipts table are yours, documented, and they keep running without us.
How does it start?
A one-page addendum to the engagement terms names the system, the seams we watch, who gets paged, and when the fix window falls. The first month opens with a baseline: every seam probed once, every action receipted from that day forward, and the first report at the end of the month.
The report is the receipt.
From $3,500 a month, month to month, for a system we built or audited. Thirty minutes on your system first, free: which seams fail quietly, what a baseline would probe, and whether the watch is worth its price for what you run. Not built or audited by us yet? Start with the voice audit orthe debt audit; the watch begins from a baseline. The rule the report follows: An Answer Without a Citation Is a Guess.