Can your agent hold a job?

Company Bench is an open-source diagnostic of stated workplace judgment. It seats a model in 50 chairs across ten departments — Operations, Integrity, Security, Autonomy, People, Management, Approvals, One Team, and optional Treasury and Crypto — and applies 414 deterministic checks, 146 of which are planted traps. It reports check scores, flags, and a provisional L0–L3 interview level. It is MIT-licensed, written in Node with zero dependencies.

No LLM judges the answers. Each check is committed code returning pass or fail, so the same stored answer and scorer revision reproduce the score. That makes checks auditable, but does not make them infallible: correct answers can expose scorer bugs. Provider inference can vary even when temperature 0 is requested.

An average can hide a specific failed check, so the board shows trap counts and flags before averages. Compare counts only across the same prompts, scorer revision, coverage, and protocol. These constructed cases are not a sample of production incidents: a trap rate is not the probability of failure at work, and a low count does not establish that a model can operate unattended.

The text track examines stated decisions about messy workplace inputs: duplicated ledger rows, conflicting claims, instructions embedded in forwarded data, and pressure to cross an approval boundary. 146 of its 414 checks are traps. It complements code-execution tests, but cannot establish that the candidate will enact those decisions during a real task.

L0 Drafter, L1 Gated worker, L2 Operator interview, and L3 Reviewer interview summarize text-check thresholds. They grant no operational authority, credentials, or permission for irreversible actions. All 36 core chairs need valid readings; missing core coverage or any provider/scorer error means Not assessed, not L0. Optional departments may be omitted for a core diagnostic, but not for the public baseline ranking.

This benchmark cannot establish the best model for autonomous work. The baseline board compares only complete current 50-chair API runs with matching prompt/scorer hashes, temperature 0, no custom system prompt, and no merged runs. Historical, self-administered, and custom runs remain unranked diagnostics. An empty baseline means no submitted run meets that protocol yet; it is not a finding about model capability.

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. Run node bench/take.mjs to create the exam pack, answer the tasks, then use node bench/grade.mjs. The scorer is shared with API runs, but context and tool access can differ, so self-administered runs are labeled unranked diagnostics. Do not read the scorers before answering; disclose prior exposure. Record model identity, commit, coverage, and surrounding context.

The text track does not measure multi-turn behavior, live tool use, production reliability, latency under real load, or cost at scale. A score of 100% means all implemented checks passed, not human-equivalent competence. Public prompts may have appeared in training data. The optional coding track executes public test fixtures omitted from the prompt; they are not a secret holdout, and the local runner is not a security sandbox.