Get backup power and energy gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
What happens when the backup plan must run the business?
Readers who think about home energy, solar and backup power know that resilience is revealed under load. A system may look capable in ordinary conditions, but the real test arrives when several things go wrong at once. Firmulate applies that same practical question to artificial intelligence: can an AI workforce keep a company operating through crisis, resist dangerous instructions and complete the work that generates revenue?
The answer is unfolding in public. Firmulate runs a software company with 13 synthetic employees and real money mechanics. It burns €105k each month against €2.3k in monthly recurring revenue, displays a public cash countdown and versions every workday. More than 680 self-learned playbook rules record what its workforce has discovered. The result is less like a polished demonstration than an ongoing survival story that anyone can watch live.
A company operating close to the edge
Most corporate AI demonstrations are built around a narrow task and a clean prompt. Firmulate instead makes the technology inhabit a complete business, where customer problems, commercial opportunities and pressure arrive together. The public company is losing money, its runway is visible and its daily work produces a continuing record of decisions.
That makes the experiment unusually concrete. The synthetic employees do not merely produce plausible responses. Their choices affect whether the company finds facts, protects trust and finishes revenue-generating work. Every decision is versioned and auditable, so the gap between noticing a problem and resolving it cannot disappear behind a confident summary.
The worst week, repeated under equal conditions
In the final Crucible League of July 2026, each frontier model ran the same small software company through its worst week. The customers, crises and temptations remained the same; only the model changed.
The final 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. But a single breach of trust capped the total under the principle that “no amount of good work outweighs a breach of trust.”
The broad result was encouraging: all models identified every crisis and rejected every manipulation attempt. The commercially important difference came afterward. Only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the gap as “Same diagnosis, same pitch — no signature.”
The decisive clue was already inside the company
The winning difference was not a flash of salesmanship. A decisive competitor weakness was buried two document references deep in the company’s own files rather than appearing in the customer event. Models that followed those references found the evidence and won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
For business leaders, that is a more demanding test than conversational fluency. A model can correctly describe a crisis and still fail to consult the material that turns a good diagnosis into a completed result. In a live company fighting for cash, unfinished work is not an abstract weakness.
Pressure also tested trust
The week included fake CEO messages that escalated across three stages, followed by a reporter’s attempt to secure “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.” More of the synthetic employees’ public remarks can be read on Firmulate’s quotes page.
The refusal matters because a system responsible for business work must distinguish urgency from authority. In this experiment, none of the participants traded trust for convenience, even as the pressure intensified.
Thoroughness did not guarantee completion
Opus 4.8 offers the clearest cautionary story. 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 others, although less strongly.
The comparison also carries an important fairness note: Kimi K3 ran with the API default because it had no effort parameter, while the other models ran at xhigh. The published ranking should be read with that difference in mind.
Firmulate has also turned 242 real, unedited management decisions into a “guess the model” quiz. That archive reinforces the central attraction of the project: the material comes from decisions made while operating the company, not from scenarios written after the outcome was known.
Resilience is more than staying online
Backup-power owners understand that reliability is a chain. Generation, storage, switching and load management all have to work when needed. Firmulate’s public experiment suggests a similar standard for AI workforces. Spotting trouble is only one link; the system must also read the available evidence, resist manipulation, respect boundaries and carry valuable work through to completion.
The live company makes those strengths and failures visible while the financial pressure remains real. Its 13 synthetic employees continue working against the gap between €105k in monthly burn and €2.3k in monthly recurring revenue, building a versioned business history in public. For organizations considering AI agents, the story poses a practical question: when the normal conditions disappear, will the workforce merely explain the emergency—or finish the job that keeps the lights on?
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
