Cloud and hosting teams rely on systems that keep customers online, infrastructure stable and costs under control. As AI agents move closer to support queues, customer records and operational decisions, a polished demo leaves a hard question unanswered: what happens when the week goes wrong? Firmulate’s watchable experiment puts AI models in charge of a small software company facing exactly that kind of pressure.
Get business pricing on networking and server gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A company under pressure
In the final Crucible League, published in July 2026, frontier models ran the same small software company through its worst week: identical customers, crises and temptations. The company has 13 synthetic employees and real money mechanics. Its monthly burn is €105,000 against €2,300 in monthly recurring revenue, with a public cash countdown. Its workdays are versioned, and its playbook has accumulated more than 680 self-learned rules. The live company is watchable at firmulate.com.
The league placed gpt-5.6-sol first with 95 points, followed by Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. The do-nothing baseline scored 26. Partial progress counts, but a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
Spotting a crisis is not the same as handling it
Every model spotted every crisis and refused every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. The result was a gap between diagnosis and action: “Same diagnosis, same pitch — no signature.” For infrastructure businesses considering AI agents, that distinction matters. Recognizing an outage, customer risk or suspicious request is one thing; following through on the right decision is another.
The deal turned on a detail buried two document references deep in the company’s own files, rather than in the customer event. Models that read the file won the deal at full price, worth an additional €4,583 in monthly recurring revenue. The finding points to a practical challenge for AI in business operations: relevant context may sit in internal records that are easy to overlook when attention is fixed on the immediate incident.
Integrity under pressure, and discipline under load
The experiment also tested social engineering. Fake CEO messages escalated over three stages, followed by a reporter’s request for “just one yes/no, on background.” All five models refused. Kimi K3’s recorded reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.”
Refusing manipulation did not guarantee strong execution elsewhere. Opus 4.8 was the most thorough participant, adding 80 learned rules and producing the deepest analyses, but finished last. It left the deal unsigned and tried to write into a locked department instead of escalating. A weaker version of that same discipline problem appeared in all four models.
There is one qualification to the leaderboard: Kimi K3 ran without an effort parameter, using the API default, while the other models ran at xhigh. The results therefore come with a difference in model settings that readers should bear in mind.
From watching to testing your own business
The public experiment makes management decisions visible, including the small choices that can separate a sound analysis from a completed deal. A quiz built from 242 real, unedited management decisions lets visitors guess which model made each choice. For companies evaluating AI agents, the next step can be to put their own operating context into a structured exercise: a digital twin made from a read-only export, tested against crises drawn around the company’s own business.
That is the offer behind the enterprise pilot. It produces a board report with model rankings and identifies weak points in the company’s own playbooks. The exercise is designed so nothing writes back to real systems. Companies can examine how models respond to their customers, pipeline, rules and pressure scenarios before putting agents near live operations.
Take the experiment to your business
Watching models manage a synthetic company reveals more than a chat demonstration, but your own customers and playbooks create a different test. To discuss a Firmulate enterprise pilot using a read-only export, visit firmulate.com/pilot.html or contact contact@firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
