Agent Behavior (website)

Agent Behavior is a format for defining, in Markdown files, the behavior that AI agents should follow across repeated interactions. It is aimed at people and agents who review agent traces and design evaluation criteria and prompts.

Each behavior spec is stored as a BEHAVIOR.md file under .agents/behaviors/ in the project repository, and is version-controlled alongside the agent code. The format defaults to free-form Markdown, but can use structures such as evidence gathering, judgment, execution, and recovery when needed.

The example spec cost-sensitive-actions requires the following behavior for actions that incur costs.

• Check the cost, credits, infrastructure impact, and alternatives.
• Determine whether the action creates a significant cost trade-off.
• Disclose the cost and request confirmation before exceeding a meaningful cost threshold.
• If the cost is unknown, investigate further or ask the user for approval.

Behavior specs are suited to behaviors that recur across multiple traces or that have a major impact on accuracy, trust, safety, cost, or user experience. They are also used to surface the agent's design direction, or to clarify criteria that would otherwise be hard to identify without separately reading prompts, tool documentation, examples, and evaluation materials.

A written spec can be used to review traces of actions such as infrastructure deployment or paid API calls, to build evaluation cases that check whether costs were disclosed, and to design prompts and tools that fill in missing behaviors.