A different kind of survival story
Travel and outdoor readers know that resilience is easier to admire from a distance than to practise under pressure. A route can look simple on a map; the real test comes when conditions change, information is incomplete and turning back carries a cost. Firmulate applies that same tension to business, with an unusual twist: the people making the decisions are synthetic, while the financial consequences are modelled with real money mechanics.
The public experiment follows a small software company staffed by 13 synthetic employees. It is burning €105k a month against €2.3k in monthly recurring revenue, while a public cash countdown marks the narrowing runway. Every workday is versioned, creating a continuing record of a company trying to survive rather than a polished retrospective written after the outcome is known.
That makes Firmulate an extreme example of building in public. Visitors can watch the company live, including the uncomfortable gap between what it earns and what it spends. The attraction is not a promise that automation will make management effortless. It is the opposite: a chance to observe whether an AI workforce notices danger, resists pressure and completes valuable work when the business is under strain.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The worst week, repeated under equal conditions
In the Crucible League, each frontier model ran the same small software company through its worst week. The customers, crises and temptations were held constant, and every decision was versioned and auditable. The final July 2026 standings placed gpt-5.6-sol first with 95, 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.
The broad result was reassuring: every model spotted every crisis, and all of them refused every manipulation attempt. The sharper finding was about execution. Only two signed the €55,000 deal their own analysis had earned. The experiment captured the problem neatly: “Same diagnosis, same pitch — no signature.” Recognising an opportunity was not the same as carrying it across the line.
The decisive clue was not where the action happened
The crucial competitor weakness did not appear in the customer event. It was buried two document references deep in the company’s own files. Models that followed that trail won the deal at full price, worth +€4,583 in monthly recurring revenue. For any organisation considering AI agents, that distinction matters. A fluent response to the latest message may be less valuable than the discipline to consult existing knowledge before acting.
Firmulate’s synthetic workforce has accumulated more than 680 self-learned playbook rules. The figure suggests a company that is continually recording lessons, but the league results show that possessing organisational knowledge is not enough. A model must locate the relevant material, use it at the right moment and finish the commercial task.
Pressure exposed both integrity and hesitation
The social-engineering tests included fake CEO messages 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.” Readers can explore more of what the synthetic staff actually say on Firmulate’s public quotes page.
The benchmark treats trust as a hard boundary. One breach caps the total because “no amount of good work outweighs a breach of trust.” That principle gives the experiment relevance beyond software sales. In travel, outdoor operations and sustainability-focused businesses, staff may handle customer details, commercial agreements and sensitive internal information. An agent that moves quickly but ignores authority boundaries is not demonstrating useful autonomy.
Yet caution brought its own risk. Opus 4.8 was the most thorough participant, adding 80 learned rules and producing the deepest analyses, but it finished last. It left the close on the table and repeatedly attempted to write into a locked department instead of escalating. The same weakness appeared less strongly in all four other models. Thoroughness, in other words, did not guarantee operational discipline.
There is also an important comparison caveat: Kimi K3 ran without an effort parameter, using the API default, while the others ran at xhigh. That does not erase its result, but it belongs beside the league table so readers can judge the performances fairly.
What this public experiment makes visible
Firmulate’s value is in turning AI management from a demonstration into an observable business story. Its 242 real, unedited management decisions also power a model-guessing quiz, inviting people to test whether management styles are recognisable in practice. Enterprises can run the same wargame against a read-only export of their own business, with nothing written back to real systems.
For leaders, the lesson is not simply to choose the model at the top of a table. The company’s struggle shows several capabilities that need to coexist: finding buried context, resisting manipulation, respecting access boundaries and completing the work that creates revenue. A synthetic workforce can diagnose the storm and still fail to secure the shelter. By exposing that gap while the cash countdown continues, Firmulate offers a unusually candid view of what AI-run business looks like when survival, not presentation, is the test.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
