The format is model- and harness-agnostic, and the specification warns clients not to inject the entire behavior corpus into runtime prompts by default.
Agent Atlas / Adjacent references
braintrustdata/agentbehavior
An open Markdown specification for versioning the recurring conduct expected from an agent, then using that answer key to review trajectories, design evals, audit prompts, and diagnose behavioral regressions.
Agent behavior specification
Full dossier
Reported + sourced
Reviewed August 2, 2026
Intended operator
Agent builders, eval authors, trace reviewers, and domain owners defining what acceptable conduct means across repeated work.
Primary surfaces
BEHAVIOR.md specification / TypeScript validator and CLI / portable authoring skill / trajectory-judging examples
Tracked repository
Teaching mechanism
The answer key is versioned separately from the test
A BEHAVIOR.md names recurring expected conduct and its failure boundary; prompts, skills, traces, rubrics, and scorers can change around it without silently becoming the source of truth.
Evidence limitThe CLI validates paths and frontmatter, not behavioral quality; useful judgments still depend on observable trajectories, calibrated fixtures, trustworthy judges, and human agreement about the intended conduct.
Inspect the source
System Map
Eight inspection boundariesProject, user, or organization behavior specs live as versioned Markdown with stable names and optional reference material.
The format does not execute tools; it can describe the evidence, decisions, visible actions, and recovery expected when an agent uses them.
A behavior can state approval or authority expectations but cannot enforce them; runtime policy and sandbox controls remain separate owners.
Whole trajectories and their recovery steps are the intended evidence surface, rather than final-answer scores alone.
Specs can define observable delegation conduct, while trace instrumentation must preserve the relevant parent, worker, and result events.
People define and revise the durable standard, calibrate it against positive, negative, lucky-correct, and outside-scope trajectories, and review disagreements at the policy, telemetry, fixture, wording, or judge layer.
BEHAVIOR.md text, structural diagnostics, event citations, true/false/not-applicable verdicts, fixture expectations, and judge-contract tests expose the chain from intended conduct to evaluation.
Learn Through This Project
8 lessons
Chapter 1
Before the First Model Request
Chapter 2
The Invisible Prompt
Chapter 3
Tool Calls: Contract, Authorization, Execution
Chapter 4
Loops, Workflows, and Recovery
Chapter 7
The Approval Trap
Chapter 8
The Human Control Loop
Chapter 9
Trust Must Make Its Case
Chapter 10
The Future Has Three Branches
Latest Reporting
1 articlePrimary Sources
Open Questions
- Can teams keep behavior specs sparse, observable, and independent enough that a correct outcome can still reveal an unacceptable process?
Evidence state describes what this Atlas profile can support today. It is not a product score and it is not evidence that uninspected capabilities are absent.