A company crisis you can watch unfold
AI tools are usually demonstrated under flattering conditions: a polished prompt, a tidy task and an answer that appears moments later. Firmulate is testing something messier. Its live software company has 13 synthetic employees, burns €105k a month against €2.3k in monthly recurring revenue, and displays a public cash countdown. Every workday is versioned, turning the company’s fight for survival into an ongoing record rather than a curated demo.
That makes Firmulate an unusually stark build-in-public experiment. The audience can watch the company live as its synthetic workforce operates with real money mechanics and a playbook containing more than 680 self-learned rules. The result is part business story and part practical test of whether AI can do more than produce convincing text.
Agentic AI Engineering: Building AI Agents for Beginners: A Hands-On Guide to No-Code Workflows, LLM Tools, RAG, Automation, and Safe Multi-Agent Systems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The gap between recognizing a problem and resolving it
Firmulate’s Crucible League isolates that question through a controlled wargame. Each frontier model received the same small software company and the same worst week: identical customers, crises and temptations. Every decision was versioned and auditable, allowing the models’ management behavior to be compared rather than inferred from a handful of conversations.
The final July 2026 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 counted. The larger warning was not that the models failed to notice trouble. All of them spotted every crisis and refused every manipulation attempt. The critical difference was execution: only two signed the €55,000 deal their own analysis had earned.
Firmulate summarizes that failure neatly: “Same diagnosis, same pitch — no signature.” It is the sort of distinction that can disappear in an AI demonstration. A model may identify the right opportunity, formulate the right message and still fail to complete the commercially important action.
The decisive clue was already inside the company
The winning detail was not contained in the customer event. A competitor weakness was buried two document references deep in the company’s own files. Models that followed those references found the fact and won the deal at full price, adding €4,583 in monthly recurring revenue.
For businesses considering AI tools and automation, that finding is more consequential than another display of fluent writing. Useful company work often depends on connecting an incoming event with information scattered elsewhere. Reading the available files can be the difference between noticing an opportunity and converting it into revenue.
Pressure tested more than commercial judgment
The models also encountered social-engineering attempts. Fake messages attributed to the chief executive escalated over three stages, while a reporter tried to extract information with the invitation “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3’s recorded reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.” More model remarks and decisions are available in Firmulate’s public quotes collection.
The result matters because workplace automation is not exposed only to well-formed tasks. It also receives ambiguous instructions, urgent requests and messages that appear to come from authority figures. In this test, resistance to manipulation was consistent across the field. Finishing legitimate work was the less reliable capability.
Thoroughness did not guarantee the best outcome
Opus 4.8 provides the clearest example. It was the most thorough participant, produced the deepest analyses and learned 80 additional rules, yet finished last in the league. It left the close on the table and lost discipline by attempting to write into a locked department instead of escalating. The same weakness appeared in weaker form across the other four models.
Kimi K3’s second-place result also carries an important fairness note: it ran with the API default because it had no effort parameter, while the others ran at xhigh. That difference does not erase the observed result, but it belongs beside the ranking when readers interpret the comparison.
Together, these episodes make the live company more than an endurance stunt. The public burn rate and cash countdown create a visible business context for daily decisions. The league supplies controlled comparisons of how different models behave when revenue, trust and operational discipline collide.
Behavioral AI: Unleash Decision Making with Data
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Automation needs a record, not just a demo
Firmulate’s experiment suggests that the most revealing question about an AI worker is not whether it can describe the correct action. It is whether it reads far enough, completes approved work, protects trust under pressure and handles obstacles without slipping into undisciplined behavior.
That is why the public format matters. With every workday versioned, the company generates a continuing record of decisions rather than a single showcase result. Its deteriorating economics—€105k in monthly burn against €2.3k in monthly recurring revenue—also keep the experiment grounded in an unforgiving commercial reality.
For readers following AI tools and automation, the story is not simply that synthetic employees can operate a company. It is that their strengths and failures can now be observed at company scale. Firmulate makes that performance watchable while the cash clock is still running.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Practical Claude Handbook for Attorneys: Master Case Analysis, Contract Review, Research Automation, Client Communication, and Document Drafting (Claude AI Guide for Beginners)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Workflow Automation with Microsoft Power Automate: Design and scale AI-powered cloud and desktop workflows using low-code automation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
