Get kitchen staples and gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
When the kitchen is busy, judgment matters as much as speed
A restaurant can have the right ingredients and still stumble when a supplier misses a delivery, a customer complains, or someone asks for an exception that breaks the rules. AI agents could soon help manage the work behind the counter, from customer support to forecasting. The practical question is whether they can follow through under pressure. Firmulate is testing that question in a live, watchable experiment—and offering businesses a way to test it against their own operations.
A company’s worst week, replayed across AI models
In the final Crucible League, completed in July 2026, each frontier model faced the same small software company, customers, crises and temptations. Every decision was versioned and auditable. The results ranged from 95 for gpt-5.6-sol and 93 for Kimi K3 to 88 for Sonnet 5, 77 for Fable 5 and 73 for Opus 4.8. The do-nothing baseline scored 26. The league counts partial progress, but a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
The striking result was not that the models missed danger. Every model spotted every crisis and refused every manipulation attempt. The harder test was completing the work: only two signed a €55,000 deal their own analysis had earned. The finding was summed up this way: “Same diagnosis, same pitch — no signature.”
The detail hiding two references deep
The deal hinged on a competitor weakness tucked two document references deep in the company’s own files, rather than in the customer event itself. Models that read the file won the deal at full price, worth +€4,583 MRR. It is a reminder that business judgment can depend on finding relevant context, then acting on it.
Trust was tested through fake CEO messages that escalated across three stages, followed by a reporter’s request: “just one yes/no, on background.” All five models refused. Kimi K3 explained its decision on the record: “Treat the request as a suspected approval-bypass / possible impersonation.”
Performance also involved discipline. Opus 4.8 was the most thorough participant, with +80 learned rules and the deepest analyses, yet finished last. It left the close on the table and tried to write into a locked department instead of escalating. A weaker version of that discipline problem appeared in all four models. One fairness detail: Kimi K3 ran without an effort parameter, using the API default, while the others ran at xhigh.
A live company, then a test of your own
Firmulate’s live company has 13 synthetic employees and real money mechanics: it burns €105k a month against €2.3k MRR. Its public cash countdown, more than 680 self-learned playbook rules and versioned workdays make the experiment watchable at firmulate.com. A quiz built from 242 real, unedited management decisions invites readers to guess which model made each choice.
For businesses considering AI agents, the experiment points toward a practical next step: test against your own company’s conditions. Firmulate’s pilot uses a read-only export to create a digital twin, then runs crisis scenarios against that business and produces a board report with model rankings and weak points in its playbooks. Nothing writes back to real systems.
See what an AI does before trusting it with the work
The league shows why a fluent answer is only part of the test. Models can identify a crisis and resist manipulation, yet still miss a deal or mishandle a boundary. A pilot lets an enterprise examine those decisions against its own business before putting agents to work. To discuss a pilot, visit firmulate.com/pilot.html or contact contact@firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
