Agent Behavior Standard Released

Basis and Braintrust have released Agent Behavior, an open standard documenting behavioral criteria for AI agents. This standard defines how agents make judgments, gather evidence, execute tasks, and recover during task completion in a single BEHAVIOR.md file, providing a single source of reference for trace review and evaluation design.

Differences from Skills and AGENTS.md

Unlike 'skills' loaded into the model at runtime or 'AGENTS.md' containing repository rules, Agent Behavior primarily targets trace review and evaluation design. Rather than being injected into runtime prompts to intentionally steer the model, it clarifies how agents should behave (normal behavioral standards) to help evaluators establish grading criteria.

Core Structure and Recommended Dimensions

The specification is located at the path .agents/behaviors/{name}/BEHAVIOR.md and consists of YAML frontmatter and free-form Markdown body. The authors recommend the following 6 dimensions to evaluate practical behavior:

• Intent: The importance of the behavior and when it applies
• Evidence: Checks and verifications before making judgments
• Decision: The reasoning and selection process
• Execution: Actual actions taken after the decision
• Recovery: Handling methods when failures or ambiguities occur
• Failure modes: Incorrect behaviors to prevent

Adoption and Validation Tools

This standard is useful for teams operating agents that perform long tasks and seeking to maintain consistency in evaluation criteria. An agentbehavior CLI tool is provided for structural checks, but it is not currently published on npm, requiring users to clone the repository and build it. The tool only validates file structure and frontmatter format, leaving the quality of the specification to human or model judgment.