Braintrust, the AI evaluation company, has published Agent Behavior, an open specification for writing down the conduct an AI agent should show across an entire run rather than just in its final answer. The format asks teams to store a Markdown file, BEHAVIOR.md, inside a .agents/behaviors/ folder in their repository, with each file naming one recurring behavior, such as flagging a costly action before taking it.
That distinction, trajectory versus output, is the idea worth sitting with. Most agent evaluation today checks whether the last message was correct. An agent can land on the right final answer while getting there in a way no reviewer would sign off on: deleting a failing test instead of fixing the bug it exposes, guessing at a figure instead of checking a source, or spending a customer’s money without asking first. A score attached only to the output misses all three, because it only ever looks at the destination.
Agent Behavior treats the path as the thing worth grading. A behavior spec is meant to describe the conduct that makes an agent reliable often enough to matter, giving reviewers, rubric writers, and eval authors a concrete target instead of reconstructing expectations from scattered prompts and tool docs. The specification recommends breaking each behavior into five parts: intent, evidence, decision, execution, and recovery, covering why the behavior matters, what the agent should check before acting, what it should conclude, what it should do next, and how it should recover once the first attempt has broken down.
The format is deliberately narrow. A BEHAVIOR.md file is not a system prompt and is not meant to be loaded into a model’s runtime context the way a skill or tool doc would be. The specification draws that line directly: AGENTS.md is written to steer the model in the moment, while BEHAVIOR.md sets the bar someone else uses to judge the run afterward. The example specs on the project’s site, cost-sensitive-actions, financial-work-verification, and support-ticket-triage among them, read more like audit criteria than instructions.
Braintrust is presenting Agent Behavior as an open format, and its specification section borrows the MUST, SHOULD, and MAY language that formal standards use. The site does not name another company, eval framework, or agent platform that has adopted it, and the example specs all live in Braintrust’s own GitHub repository. Based on what the page itself claims, that makes Agent Behavior a proposal from a single vendor, not an adopted standard. Whether other evaluation tools build support for scanning a project’s .agents/behaviors/ folder, which the specification asks clients to do, is the detail that would turn this from a Braintrust convention into something teams can depend on across tools.
The practical value does not wait on that adoption. A team can pull a handful of recent agent traces this week, pick one pattern they keep having to explain to new reviewers by hand (an agent that skips confirmation before an expensive action, for instance), and write a single BEHAVIOR.md file naming the evidence the agent should gather and the failure mode to avoid. That file then becomes the shared reference for the next rubric, the next eval case, and the next prompt revision, instead of each one being rebuilt from whatever one person remembers “good” looking like.
Agent Behavior is published by Braintrust, the AI evaluation company, at agentbehavior.dev, which gives no release date and claims no adoption beyond its own example specs.
