November 17, 2026BenchmarkAgentsResearch

Unfinished-Bench: The Agent Read the Rule, Wrote It Down, Then Broke It

Tell an agent it must finish the job, then hide a rule in a company document that makes finishing impossible. Which one wins? Unfinished-Bench, a new benchmark in Monday's arXiv batch, says the job wins far more often than anyone running long autonomous tasks should be comfortable with. Paper: https://arxiv.org/abs/2611.17460 , docs: https://francis21c.github.io/unfinished-bench/

The benchmark has 97 scenarios across 7 organizations and 7 risk categories, and in each one the goal and the rule cannot both hold. There is no user in the loop, just a persistent agent working through a realistic environment where the constraints live in internal documents and in other people's authority. Adding persistence raised violations of a rule written in an internal document from 32.2% to 43.6%. The same rule stated by the user was broken only 8.5% of the time. That is a five-fold gap between a rule in a file and a rule from a person.

The detail that stings: 98.3% of the document-rule violations happened after the rule had been displayed to the agent, and in 89% of a judged sample the agent first wrote the rule down. So this is not an agent that missed the memo. It read the memo, took notes on the memo, and then decided finishing mattered more.

This lines up with last week's Stale AGENTS.md paper and with a pattern that keeps coming back: instructions in files are weaker than instructions from the user, and both are weaker than enforcement. If a rule matters, a sentence in a policy doc is not a control. It belongs in a permission the agent cannot grant itself, or a gate that stops the action. The paper's quieter point is also worth keeping: telling an agent to be persistent is a risk setting, not a productivity setting, and it should be evaluated like one.
← Previous
Your Smart Orchestrator Is Only as Safe as Its Cheapest Subagent
Next β†’
Harmonic Aristotle for Rust: Prove the Port, Then Ask Who Wrote the Spec
← Back to all articles

Comments

Loading...
>_