Inspect AI (ukgovernmentbeis/inspect_ai) is a framework for evaluating large language models. According to the release notes published on GitHub's 0.3.282 tag, this version was released on October 19, 2026 (UTC), and contains a change that separates tool state for each agent running in parallel on the same sample.
Evaluation is the procedure of giving a model tasks and measuring quality by comparing the results against criteria. An evaluation framework is a tool that defines those tasks and scoring criteria in code, runs the model repeatedly, and compares the results. Inspect AI is a publicly available repository on GitHub, and it shows 3057 stars. The target of this change is the evaluation style that attaches multiple agents in parallel to the same sample. Separating tool state per agent keeps one agent's tool usage history from being mixed with another agent's. The same release also includes a conditional save and lock helper for sharded execution (running work split into several chunks), along with fixes related to cancelled evaluations and tool call and usage aggregation.
However, this release description contains no performance figures and no comparison of scoring results. It does not explain whether separating tool state changes existing evaluation results, nor under what conditions the conditional save and lock helper for sharded execution operates. The change applies directly to parallel multi-agent evaluations, but its impact on single-agent evaluation results cannot be known from this release description alone.