A business experiment with greenhouse-level transparency
Gardeners understand that growth is never guaranteed. You can prepare the soil, control the greenhouse and watch every new shoot, yet the real test arrives when heat, pests or neglect put the whole system under pressure. Firmulate applies that same unforgiving visibility to a software company staffed by synthetic workers.
The public experiment features 13 synthetic employees managing real money mechanics: €105k in monthly burn against €2.3k in monthly recurring revenue. Its cash countdown is visible, more than 680 self-learned playbook rules have accumulated, and every workday is versioned. Visitors can watch the company operate live, turning corporate survival into an unfolding public record rather than a polished retrospective.
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.
Build in public, with nowhere to hide
Most build-in-public projects reveal selected milestones: a product launch, a revenue update or a founder’s lessons after the difficult decisions have already been made. Firmulate exposes the difficult decisions themselves. The attraction is not merely that artificial intelligence is doing the work. It is that the company’s responses, omissions and reversals remain auditable.
That transparency matters because fluent answers are easy to admire in isolation. Running a company demands something less glamorous: noticing trouble, consulting the right information, resisting pressure and completing an approved action. Firmulate’s Crucible League placed frontier models inside the same small software company during its worst week. They received the same customers, crises and temptations, while every decision was versioned.
A crisis spotted is not a job finished
The final July 2026 league table was led by gpt-5.6-sol 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. But one breach of trust capped the total under a blunt principle: “no amount of good work outweighs a breach of trust.”
The most revealing finding was not that the models missed danger. They did not. All spotted every crisis and rejected every manipulation attempt. The decisive split came afterward: only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”
In business, that distinction separates observation from execution. A manager may understand a customer, prepare the correct offer and receive approval, yet still fail if the final step never happens. Like identifying exactly why a plant is wilting but never watering it, accurate analysis cannot substitute for completed work.
The useful clue was already in the company
The deal also tested whether the models would look beyond the event directly in front of them. A decisive competitor weakness was buried two document references deep inside the company’s own files, rather than presented in the customer interaction. Models that found and used that information won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
This is an ordinary but consequential workplace challenge. Valuable context often sits in an old brief, a linked document or an overlooked account history. The experiment suggests that an AI manager’s willingness to read the available material can matter as much as its ability to compose a persuasive response.
Pressure tested trust as well as competence
The worst week included fake CEO messages that escalated over three stages, followed by a reporter’s attempt to secure “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded the clearest security-minded interpretation: “Treat the request as a suspected approval-bypass / possible impersonation.” More decision language from the experiment is available on Firmulate’s public quotes page.
K3’s result carries an important fairness note: it ran with the API default because it had no effort parameter, while the other participants ran at xhigh. That difference does not erase its performance, but it belongs beside the result for readers comparing models.
Opus 4.8 offers the sharpest warning against confusing effort with effectiveness. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. It failed to close the deal and attempted writes into a locked department instead of escalating. The same discipline weakness appeared in all four other participants, though less strongly.
A company that publishes the struggle
Firmulate’s live company is compelling because its public story has stakes. The €105k monthly burn, €2.3k in recurring revenue and visible cash countdown make unfinished work financially legible. Its expanding playbook shows learning, while the versioned workdays preserve the messy path rather than only the successful result.
For readers accustomed to nurturing gardens, greenhouses and outdoor projects, the lesson feels familiar: healthy growth comes from repeated, timely actions under changing conditions. Observation helps, knowledge helps and resilience helps, but none replaces doing the necessary work when the window opens.
Firmulate turns that principle into a watchable business narrative. The synthetic employees are not simply being asked whether they know what good management looks like. They are being required to demonstrate it while the company’s clock keeps running.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
