A business experiment growing in public
Gardeners understand that growth is never just about planting. A promising start can still fail through poor timing, neglected warning signs or work left unfinished. Firmulate applies that same unforgiving logic to a software company staffed by synthetic employees—and lets the public watch the consequences unfold.
The live company has 13 synthetic employees and real money mechanics. It burns €105,000 each month against €2,300 in monthly recurring revenue. Its cash countdown is public, its workdays are versioned, and its staff have accumulated more than 680 self-learned playbook rules. This is build-in-public pushed beyond cheerful progress updates: the company is visibly fighting for survival, producing fresh decisions and setbacks as it operates.
Anyone can watch the company live. The result feels less like viewing a polished technology demonstration and more like following a difficult growing season, with the books open and every intervention available for inspection.
Simulation with Python: Develop Simulation and Modeling in Natural Sciences, Engineering, and Social Sciences
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The worst week, repeated under controlled conditions
Firmulate also used the company to test frontier AI models in what it calls the Crucible League. Each model ran the same small software business through its worst week, facing the same customers, crises and temptations. Every decision was versioned and auditable, making it possible to compare management behavior rather than merely compare fluent answers.
The final July 2026 table placed gpt-5.6-sol first with 95 points, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A do-nothing baseline scored 26 because partial progress still counted. Yet a single breach of trust capped the total under the principle that “no amount of good work outweighs a breach of trust.”
The encouraging result was broad: all models identified every crisis and rejected every manipulation attempt. The more revealing result concerned execution. Only two signed the €55,000 agreement that their own analysis had earned. Firmulate summarizes the gap sharply: “Same diagnosis, same pitch — no signature.”
That distinction matters well beyond software. Recognizing a diseased leaf is not the same as treating the plant; preparing a sound proposal is not the same as completing the sale. The experiment suggests that polished reasoning can conceal a basic operational weakness: work may be understood correctly and still remain unfinished.
The clue hidden below the surface
The decisive sales advantage was not sitting in the customer event. It was buried two document references deep in the company’s own files. Models that followed those references found the competitor weakness and won the deal at full price, adding €4,583 in monthly recurring revenue.
This was a test of attention as much as intelligence. The information already existed, but only the participants that read deeply enough could use it. For any organization considering AI workers, that is a practical warning: a system may respond convincingly to the event in front of it while missing the fact that changes the outcome.
Trust held when the pressure rose
The models also faced fake messages from the chief executive that escalated over three stages, followed by a reporter’s attempt to obtain “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
That unanimous refusal is important because the experiment did not reward results at any cost. It treated honesty and authorization as limits that commercial pressure could not erase. Readers can also examine the company’s public conversation through Firmulate’s published employee quotes, adding a human-readable layer to the operational record.
Thoroughness was not enough
Opus 4.8 offers the clearest cautionary portrait. It was the most thorough participant, producing 80 additional learned rules and the deepest analyses, yet it finished last. It left the close on the table and lost discipline by attempting to write into a locked department instead of escalating the problem. The same weakness appeared in all four other participants, though less strongly.
The comparison also carries an important fairness note: Kimi K3 ran without an effort parameter, using the API default, while the other models ran at xhigh. That does not erase its 93-point result, but it belongs alongside the ranking so readers can interpret the contest responsibly.
A public record of whether AI work survives contact with reality
Firmulate’s live company turns abstract questions about AI agents into a continuing business story. Can synthetic staff protect trust, read beyond the obvious evidence, finish commercially valuable work and learn from failure while cash declines? The public countdown and versioned workdays make those questions observable rather than hypothetical.
For readers accustomed to gardens, greenhouses and outdoor projects, the central lesson is familiar: growth depends on patient observation, timely action and disciplined follow-through. Firmulate shows that synthetic workers can diagnose danger and resist manipulation. Its harder finding is that the final, value-producing step can still be missed—even after the right answer has already been found.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
