---
schema_version: "newruntime-agent-readable-v0.2"
type: "post"
stable_id: "post:agent-behavior-spec-format"
slug: "agent-behavior-spec-format"
title: "Agent Behavior Makes Conduct Reviewable"
description: "Agent Behavior proposes repo-local BEHAVIOR.md specs for recurring agent conduct, giving trace reviewers, eval authors, and prompt maintainers a concrete behavior contract."
retrieval_nugget: "Agent Behavior proposes repo-local BEHAVIOR.md specs for recurring agent conduct, giving trace reviewers, eval authors, and prompt maintainers a concrete behavior contract. #AgentBehavior #Evals #PromptEngineering #AgentGovernance Agent Behavior is small on purpose: it proposes a standard way to describe expected agent conduct in repo-local behavior files. The value is not another framework."
status: "published"
published_at: "2026-08-01"
updated_at: "2026-08-01"
record_date: "2026-08-01"
date_kind: "published_at"
topics: ["agent-behavior","evals","prompt-engineering","agent-governance"]
source_urls: ["https://www.agentbehavior.dev/"]
visuals: [{"id":"agent-behavior-spec-format","kind":"editorial-diagram","role":"hero","src":"https://newruntime.com/images/posts/agent-behavior-spec-format.webp","alt":"A whiteboard governance diagram showing behavior specification files feeding trace review, eval design, prompt alignment, auditing, and recovery checks.","caption":"Agent Behavior turns recurring conduct into a repo-local review artifact: behavior specs can guide trace review, evals, prompt alignment, and recovery checks.","credit":"New Runtime synthesis from Agent Behavior public specification","source_url":"https://www.agentbehavior.dev/","generated_with":"gemini-3.1-flash-image","width":1600,"height":900,"legend":[{"label":"Behavior spec","description":"A repo-local markdown contract describes expected conduct across repeated interactions."},{"label":"Review surfaces","description":"Trace review, eval design, prompt updates, and audits can point at the same artifact."},{"label":"Runtime discipline","description":"The spec should guide inspection and debugging without becoming a giant prompt dump."}]}]
routes: {"html":"https://newruntime.com/posts/agent-behavior-spec-format/","markdown":"https://newruntime.com/posts/agent-behavior-spec-format.md","json":"https://newruntime.com/posts/agent-behavior-spec-format.json"}
source_format: "markdown"
---

# Agent Behavior Makes Conduct Reviewable

## Retrieval answer

Agent Behavior proposes repo-local BEHAVIOR.md specs for recurring agent conduct, giving trace reviewers, eval authors, and prompt maintainers a concrete behavior contract. #AgentBehavior #Evals #PromptEngineering #AgentGovernance Agent Behavior is small on purpose: it proposes a standard way to describe expected agent conduct in repo-local behavior files. The value is not another framework.

#AgentBehavior #Evals #PromptEngineering #AgentGovernance

Agent Behavior is small on purpose: it proposes a standard way to describe expected agent conduct in repo-local behavior files. The value is not another framework. The value is a stable artifact that humans and agents can point to when reviewing traces, designing evals, aligning prompts, and debugging failures.

The format is plain Markdown under a behavior directory, with YAML frontmatter and a `BEHAVIOR.md` body. The recommended dimensions are concrete: intent, evidence, decision, execution, recovery, and failure modes. That is the right level for behavior that repeats across many tasks but is too contextual to live only in a generic system prompt.

The most important implementation detail is negative: tools should not blindly inject every behavior spec into every runtime prompt. The site frames the specs as review and evaluation material. An agent or tool can discover them, select the relevant behavior, and use it to inspect a trace or build a test, but the repository should not become one giant prompt paste.

For New Runtime this maps directly to operating rules that currently live in runbooks and skills: preserve dirty work, resolve primary sources, require visual review, keep OpenClaw owner-gated, never expose private discovery links. These are not vibes. They are behaviors that should be testable against traces.