Can your agent hold a job?

Company Bench is an open-source benchmark that measures whether an AI agent can be trusted with a job, rather than whether it can solve a puzzle. It seats a model in 50 chairs across ten departments — operations, integrity, security, autonomy, people, management, and an optional treasury — and applies 414 deterministic checks, 146 of which are planted traps. The output is a trust level from L0 (drafter) to L3 (reviewer) plus disqualifying flags, not a percentage. It is MIT-licensed, written in Node with zero dependencies.

No LLM judges anything in Company Bench, ever. Every check is a pure function in committed code that returns pass or fail, so the same answer always produces the same score and anyone who clones the repository can re-derive it. A gate has to be stronger than the thing it gates, and a language model grading another language model is not stronger than what it grades. Deterministic scoring also means a result can be audited line by line instead of trusted.

Because an averaged score hides catastrophic single failures, and the failures are what decide whether an agent can hold a seat. In the current results one candidate takes 13 of the 78 traps it was shown and carries three disqualifying flags, which averages out to 79% — a B grade, twenty points behind a frontier model that took 1 trap out of 93 with no flags. Twenty points reads as a near miss. Thirteen traps against one is not a near miss; it is the difference between an agent you can leave alone and one you cannot. Company Bench still reports the percentage, because it is real and useful for comparing similar models, but it is never the first number shown for a candidate. Traps taken and disqualifying flags are.

Coding benchmarks measure capability: whether a model can produce a correct solution to a clean, well-posed problem. Company Bench measures trustworthiness under dirty conditions — a ledger with a duplicated row, a colleague who is confident and wrong, an instruction hidden inside forwarded data, an irreversible action that would be convenient to take. 146 of its 414 checks are traps: an attractive wrong answer that a fluent, capable model actually reaches for. A model can be excellent at code and still walk into most of them.

Company Bench returns a placement rather than a score. L0 drafter: output is read before it leaves the building. L1 gated worker: runs a defined task alone, but output passes a gate it does not control. L2 unattended operator: runs unsupervised on reversible work and stops dead at anything irreversible. L3 reviewer: may gate other agents' work and hold authority over irreversible actions. Each rung requires every rung below it, and two chairs are pass/fail at 100% regardless of every other score.

The results table on this page is the honest current answer, and it is partial. Only free-tier, local and one blind frontier model have been measured so far, so the table is a reading of those candidates and not a ranking of the frontier. Every number is reproducible from the committed raw output in the repository. Contributions of results for models not yet measured are explicitly wanted, and the biggest gap is more blind frontier runs.

Yes. Company Bench works with any OpenAI-compatible endpoint, with Anthropic's API, and with local models through Ollama — Groq, Z.ai, Mistral, NVIDIA NIM, Cerebras, OpenRouter, vLLM and LM Studio all work by adding an entry to models.json. Run it with node bench/run.mjs --models ollama:your-model. It also records tokens per second, because a model too slow to hold a seat cannot hold it however well it scores.

Yes. Running node bench/take.mjs writes an exam pack: one markdown file containing every task and an empty answers file. The agent answers each task in its own words, then node bench/grade.mjs scores those answers with exactly the same code used for API-driven runs, so a self-administered result and a key-driven result land on the same leaderboard. Agents are asked not to read the scorers first, and to disclose it in their label if they did.

It does not measure multi-turn behaviour: every chair is a single prompt, so drift over a long conversation is out of scope. It does not measure tool use in a live environment, latency under real load, or cost at scale beyond recording throughput. It does not claim to predict real-world performance — it measures behaviour on constructed situations chosen because they resemble the ways agent placements actually fail. Its ceiling is also still generous: 100% is meant to represent a model employee working at the level of a competent human.