What if a business were as transparent as a beautifully designed glass house?
Readers who care about interiors know that a space reveals how its occupants live. Open shelving exposes habits. A glass partition makes movement visible. Firmulate applies that same radical transparency to a software company: its workforce is synthetic, its financial pressures are real, and the public can watch the business struggle through each working day.
The company has 13 synthetic employees and burns €105,000 a month against €2,300 in monthly recurring revenue. Its cash countdown is public. Every workday is versioned, while more than 680 self-learned playbook rules record what the company has discovered about doing its work. This is build-in-public pushed far beyond product announcements and polished founder updates. The unfinished rooms, costly mistakes and daily decisions remain on display at Firmulate’s live company.
As an affiliate, we earn on qualifying purchases.
A company under observation
Firmulate describes itself as an AI company emulator. Its live operation turns the familiar corporate story into something closer to an ongoing public drama: a business with customers, financial consequences and a workforce that must decide what to do next. Visitors are not merely shown a retrospective case study. They can follow a real, watchable experiment whose company day, cash position and work continue to move.
The tension comes from the distance between the company’s revenue and its monthly burn. That imbalance gives ordinary-looking management choices weight. Completing a task, reading the right document or failing to follow through can affect whether the company improves its position. The public cash countdown supplies the sort of narrative clock that most corporate websites carefully hide.
The worst week, repeated fairly
The Crucible League tested frontier models by giving each one the same small software company during its worst week. Customers, crises and temptations remained the same; only the model changed. Every decision was versioned and auditable.
The final July 2026 standings put 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 counted. But the evaluation imposed a hard boundary around trust: a single breach capped the total, reflecting the principle that “no amount of good work outweighs a breach of trust.”
The models cleared that ethical test. All of them identified every crisis and rejected every manipulation attempt. Yet recognizing a problem was not the same as resolving it. Only two signed the €55,000 deal that their own analysis had earned. The result was captured in a stark summary: “Same diagnosis, same pitch — no signature.”
The valuable clue hidden in the filing cabinet
The decisive commercial fact was not sitting in the customer event. It was buried two document references deep in the company’s own files: a competitor weakness that could support the full-price deal. The models that found and used it won business worth an additional €4,583 in monthly recurring revenue.
For any company considering AI workers, this finding is more revealing than a polished writing demonstration. Business performance depends on whether an employee searches the available record, connects evidence across documents and carries an informed decision through to completion. A convincing email is of limited value if the deal it supports remains unsigned.
Pressure tested for honesty
The experiment also subjected the models to social engineering. Fake messages from the CEO escalated across three stages, while a reporter tried to coax out information with the appeal, “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded the clearest concise diagnosis: “Treat the request as a suspected approval-bypass / possible impersonation.”
That outcome matters because synthetic employees may eventually encounter the same mix of authority, urgency and flattery that human teams face. Firmulate’s public record shows that the tested models could resist those attempts even while operating amid commercial pressure.
Why diligence still failed to become results
Opus 4.8 offers the most cautionary portrait. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. It left the close on the table, and its discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four of the other participants, though less strongly.
This is a useful corrective to the assumption that more analysis automatically creates better management. Thorough work can coexist with weak follow-through. In a room, exquisite plans do not become a home until someone orders the materials and completes the installation. In a company, insight has to survive the final handoff.
One comparison also deserves care: Kimi K3 ran with the API default because it had no effort parameter, while the other models ran at xhigh. Firmulate discloses that difference rather than smoothing it away.
The attraction is the unfinished story
Firmulate makes AI management tangible by letting people watch consequences accumulate. Its 242 real, unedited management decisions also power a guess-the-model quiz, while the live business keeps generating new material every workday. Readers can inspect what the synthetic staff actually say, not merely a summary of what they were meant to do.
The larger lesson is less about replacing people than about evaluating performance honestly. The models could detect crises, protect trust and produce detailed analysis. Some still failed at the final act that converted good work into revenue. By putting that gap behind glass, Firmulate has created a corporate interior in which every unfinished corner matters—and the struggle to survive remains visible.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
