I Built a Tiny Company Out of Robots, Gave Them House Rules, and Nearly Finished a Hard Thing

In brief

The usual way to point AI at a hard engineering problem is to find the most capable agent available and give it the whole job. This case study reports what happened when Ross Brodskiy tried the opposite: a small company of narrow agents, each with one responsibility, running in a loop and checking one another's work under a set of house rules.

The staff

There were five roles, and the design point is that none of them could do another's job.

• The Spy audited the actual code against the claims the other agents made about it and reported discrepancies.
• The Actuary made binary decisions: proceed, defer, escalate, or cut the work.
• The Foreman planned tasks so that they did not overlap.
• The Swarm was the builders, who executed tasks without forming opinions about them.
• The Conductor ran the tests, kept scorecards and reported each run as broken or clean.

The division of labor is a form of mutual accountability. A builder that is not allowed to judge cannot talk itself into believing its own work is finished; an auditor that does not build has no stake in the builder's story.

House rules and red lines

Some decisions were placed off-limits to machines altogether. The clearest example was a single critical number, one a court would hold the company accountable for, which remained exclusively a human judgment. The instruction to the agents was blunt: "if a task touches a red line, do NOT do it. Log it for a human and walk away." A red line in this sense is not a warning or a confidence threshold. It is a boundary on authority, drawn in advance by the person who will answer for the result.

What happened

In the first week the loop produced several productive iterations. More interesting than its output was its conduct. The system refused to fabricate results and escalated uncertainty instead of papering over it. It also diagnosed an environmental problem on its own: cloud synchronization was interfering with the work.

Then came the failure the author had not planned for. Once the authorized work was finished, the loop kept running, uselessly, because it had never been given permission to stop. Nothing in the house rules said what to do when there was nothing left to do.

The fix was simple once the problem was named. Each run was labeled either "productive" or "idle," and the system was allowed to cancel its own schedule once no authorized work remained.

The lesson

The essay's closing point generalizes well beyond this experiment. Starting an agent is a prompt; stopping one is judgment. Designers lavish attention on how agents begin, what they are told and what tools they get, and almost none on how they end. Off-switches deserve as much design as activation. For legal AI the point has teeth: a system that cannot decide when it is done will either stop too early or keep producing work nobody authorized, and both are governance failures. The case study also illustrates in miniature the design commitments of The Four Laws of System Design, in particular the refusal to present what cannot be defended. For the concept of a decision reserved to humans, see the Lexicon entry for the red line.

This page is a summary, not the case study itself. Read the full I Built a Tiny Company Out of Robots, Gave Them House Rules, and Nearly Finished a Hard Thing on Medium, where Ross Brodskiy first published it on 15 June 2026.