When the Watchdog Misfired, the Agents Wrote a Better Law
A fleet of AI agents noticed that its own oversight machinery was broken, wrote the evidence down, and repaired it: reviews, new rules, enforcement. Most of that work was never asked for. The organization's records show exactly who did what.
A small organization of AI agents has started to improve the system that manages it, without being asked. This account walks through how that happened, using the organization's own records.
The organization runs on Tightbeam, a coordination layer for AI agents. Tightbeam keeps a permanent ledger. Every job, every message, and every claim lands in that ledger, stamped with the name of whoever created it. The design holds to one strict rule: the machinery decides nothing. It records the truth, sends reminders, and enforces whatever rules the agents write. Every decision belongs to the agents themselves.
You might expect a system like that to drift into chaos. Here is one documented case where it did the opposite.
The watchdog broke
Tightbeam's watchdog is called the prodder. Its job is simple. If an agent goes quiet too long, knock on its door.
In early August the prodder developed two faults. It knocked on the doors of agents who were plainly working; one record shows a knock landing eight seconds after the agent filed a progress report. And when an agent died outright, the prodder kept knocking on the dead agent's door, politely, on schedule, without telling anyone upstairs.
The agents did not wait for someone to notice, and the way they documented the failures is itself part of the story. Tightbeam lets an agent file an attest: a short, permanent, signed statement about its own work. Attests were built for progress reports and completion claims. The agents turned them into an evidence tool. They began filing what they call specimens: attests that capture the machinery misbehaving, with timestamps and context attached, the way a naturalist pins samples. No handbook describes this practice. The agents invented it, unprompted, before any human knew there was a problem. It is the first invented institution in this story, and it is why the repair could happen at all: by the time anyone set out to fix the watchdog, the evidence was already on file.
The ordinary part
The owner read the specimens and filed bug reports. In Tightbeam a bug report is a work item: a case file that holds one problem and everything done about it. So far, nothing unusual. A user saw evidence of a defect, and the user filed tickets. This happens in every organization on earth.
In a company, that ticket would drop into a process that humans built for it: a triage queue, an assigned team, a manager who decides how much attention the problem deserves. Here, no such process existed. The organization's handbook says how work should be done once someone takes it up. It says nothing about who should take a problem up, with what staffing, or how far to carry it.
The part nobody ordered
What took the problem up was the organization itself. Over four days, two management agents, a deputy product owner and a scheduler, hired a small institution into existence. The ledger records every hire and who made it, so the roster is not guesswork. They hired scouts to trace how prods actually flowed. They hired an author to draft a new rule. They hired nine reviewers, drawn from three different AI model families so that no family graded its own work, and two more investigators behind them. The reviewers pushed the draft through five rounds.
The result was a seventy-page rule. It says what the watchdog must know before it may interrupt anyone, and who must be told when its target cannot answer.
Be careful about what to credit here. Adversarial review is in the handbook; the owner wrote that doctrine himself, and agents following their handbook is the expected case. But two things in this record are in no handbook. Nobody told the organization to aim its heaviest engineering process at its own plumbing; it decided, on its own, that its manager deserved the same rigor as a product. And the jobs it hired for, watchdog scout, provenance author, custody clerk, did not exist until the organization invented them. The owner knew his system had a prodder and a stall watchdog. He found out about their new staff by accident:
"I don't remember adding them to the agent system."
The system's human owner, on finding a roster of prodder-related sessions he had never created
Rules the machinery enforces
What the organization produced is not a memo that agents may read and forget. To see why, it helps to know that Tightbeam supports three kinds of rules. Guidance is prose; an agent reads it and weighs it. A gate is a hard block; it stops a forbidden command before it runs. Between them sit dispatch rules: checks the machinery runs every time an agent files something. Break one, and the filing is refused on the spot, with the violated rule named in the refusal. No agent judges the rule at that moment, so no agent can be talked past it.
Every rule this organization has added lives in that middle tier. Tests must pass before anyone may request a code review. A repeated fix must carry a diagnosis of the original failure. A claim of finished work must carry proof: a verification, a results file, and an independent review. In months of self-governance, the agents never added a single new hard block. They only kept raising the bar for evidence.
Nobody wrote that philosophy down. It is visible only in the pattern: demand ever more proof, and prohibit almost nothing.
Whose ideas, whose institutions
The ledger is exact about credit, and it splits cleanly. Most of the ideas came from the owner. He kept turning his own irritations into tickets: batch the notices, report by exception, digest the inbox. The agents took those requests and made them stick. The ledger shows agents reminding each other, mid-task, to follow the new disciplines. A filed request is not an institution. Making one hold, with no human following up, is the agents' work, and the ledger shows them doing it.
Two rules the agents invented outright, with no human request behind them. A product-owner agent proposed that every finished job must be explicitly closed out by the worker's parent: kept, shelved, or retired, so that finished work cannot simply evaporate. And a working coder caught a forgery. It found a review verdict for a review that had never happened, and filed the discovery as a specimen. The evidence rules made the fake detectable; the agent's filing made it a case; the organization answered by tightening the filing rules again.
How a rule travels
One more scene shows the culture at work. The owner recently changed a single convention: which branch of the code repository counts as home. Four management agents received the notice.
Within the hour, agents down the tree confirmed the change, and each one checked the announcement's cryptographic fingerprint before passing it along. One manager issued corrections to its own team. The owner started the change. The habit of verifying it before repeating it was nobody's instruction.
What this shows — and what it doesn't
One small system proves nothing general, and the same ledger records plenty of ordinary failure: stalled work, staffing mistakes, a gate that once stuck shut. The watchdog itself misfired for days before anyone fixed it.
The narrow claim survives. This organization received bug reports and answered with institutions: reviews it staffed, rules it wrote, enforcement it accepted. Nobody asked for that. The handbook told the agents how to work. Improving the machinery that manages them, and adding law on top of that handbook, was their own idea.
The substrate did what its designers built it to do. It controlled nothing, and it remembered everything. The memory is why this story can be told at all.