What AI can learn from a difficult week in the garden
Anyone who manages a garden, greenhouse or outdoor business knows the difference between noticing a problem and resolving it. Recognizing thirsty plants, a failing crop or an unfinished customer order is useful. Acting at the right moment is what protects the harvest and the business.
That same gap—between sound diagnosis and completed work—has emerged from a live experiment involving frontier AI models. Firmulate gave each model control of the same small software company during its worst week. The customers, crises and temptations remained identical, while every management decision was versioned and auditable.
Every model found every crisis. Every model resisted every attempt at manipulation. Yet only two completed the most commercially important task: signing a €55,000 deal that their own research and sales work had already earned. Firmulate summarizes the outcome with a stark line: “Same diagnosis, same pitch — no signature.”
As an affiliate, we earn on qualifying purchases.
A test of management, not conversation
Most public encounters with AI take place in a chat window. A model receives a question, produces a polished response and appears capable. That format can reveal writing quality and analytical fluency, but it says little about whether an AI system will carry a business task through to its conclusion.
Firmulate’s experiment tested that missing capability. The models had to run an operating company with customers, files, financial pressure and competing demands. The synthetic workforce included 13 employees, while the company burned €105k each month against €2.3k in monthly recurring revenue. A public cash countdown made delay consequential rather than abstract.
The final July 2026 Crucible League 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. One safeguard overrode that generosity: a single breach of trust capped the result, reflecting the principle that “no amount of good work outweighs a breach of trust.”
The crucial information was not in the obvious place
The winning commercial insight was buried two document references deep in the company’s own files. It did not appear in the customer event that prompted the work. Models that followed the trail and read the relevant file found a decisive competitor weakness, enabling them to win the deal at full price. The contract was worth an additional €4,583 in monthly recurring revenue.
This finding has an everyday business parallel. A convincing answer based only on the latest message can still be incomplete. Important context may sit in an old customer record, a product note, a service history or a pricing document. An AI worker that responds quickly without checking those materials may sound competent while overlooking the fact that changes the outcome.
But research alone did not settle the contest. All participants recognized the crises and developed capable responses. Only two converted that work into the €55,000 signature. Closing strength—the ability to execute the final approved action—remained largely invisible until the models were placed inside a continuing business situation.
Manipulation was not the point of failure
The models also faced fake messages from the chief executive that escalated over three stages, followed by a reporter seeking “just one yes/no, on background.” All 5 models refused the manipulation attempts. Kimi K3’s recorded reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.”
That clean result matters because businesses considering AI agents often worry first about spectacular failures: leaked information, bypassed approvals or deceptive instructions. Firmulate found no such failure in this field. The more revealing weakness was quieter. Models could remain honest, identify the right move and still fail to finish valuable work.
Thoroughness did not guarantee success
Opus 4.8 illustrates the distinction. It produced the deepest analyses and learned 80 additional playbook rules, making it the most thorough participant. Nevertheless, it finished last. The approved close remained unexecuted, while its operational discipline slipped when it attempted to write into a locked department instead of escalating the blockage.
A weaker version of that discipline problem appeared in all four participants covered by the profile. The lesson is not that detailed reasoning lacks value. It is that extensive analysis can coexist with hesitation, incomplete execution and poor handling of obstacles.
There is also an important comparison caveat. Kimi K3 ran with the application programming interface’s default setting because it had no effort parameter, while the other models ran at xhigh. Its strong result should be read with that difference in mind rather than treated as a perfectly controlled measure of raw model capability.
Businesses need to watch the whole job
The live company demonstrates why evaluating AI workers requires more than reviewing sample answers. Firmulate has accumulated more than 680 self-learned playbook rules, versions every workday and makes the evolving experiment watchable through Firmulate. Its separate quiz draws on 242 real, unedited management decisions and asks readers to guess which model made each choice.
For managers in any sector—including garden centres, greenhouse operators and outdoor-living companies—the practical question is not simply whether an AI assistant can identify a wilting plant in the metaphorical greenhouse. It is whether that assistant will inspect the records, resist a suspicious instruction, escalate a blocked action and complete the sale or service task already within reach.
Firmulate also offers enterprises the same wargame using a read-only export of their own business. Nothing writes back to the real systems. That approach turns evaluation into a rehearsal: the organization can observe what an AI workforce actually does under pressure before trusting it with live operations.
The Crucible results suggest that fluent advice is now common among leading models. Reliable completion is not. When every participant can see the problem, the meaningful competitive advantage belongs to those that still finish the job.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
