Agent OS is a public knowledge bundle of concise operating principles, policies, playbooks, practices, patterns, and dated observations for agentic work.
The repository uses Markdown as its canonical format. Content under kb/
targets Open Knowledge Format 0.1.
Start with the knowledge index. Treat policies and principles as stable guidance, playbooks as repeatable procedures, and observations as dated evidence that may need retesting.
Agents should use the task-scoped usage guide to load only the concepts relevant to the current task.
For controller-worker execution, use the Fast Goal and Deep Goal framework: Fast is the default, and Deep is reserved for genuinely risky or complex work.
Learn from Agent OS before continuing.
Agent OS source: <AGENT_OS_ROOT, configured workspace, or public Agent OS URL>
Active project: <active-project-root or repository context>
Task: <task>
Learn: Locate the configured clone through AGENT_OS_ROOT or workspace discovery.
If it is unavailable, use the public Agent OS knowledge base. Read Agent OS
AGENTS.md, README.md, and kb/index.md first.
Apply: Choose the matching task route and load only one to three relevant
concepts. Briefly state what applies, what does not apply, and any conflict. The active
project's instructions, SPEC, approvals, and safety policy remain authoritative.
Apply the useful guidance to planning and execution.
Contribute: At closeout, name the Agent OS concepts used and suggest reusable
learning as a sanitized candidate. Do not publish private runtime truth or
modify Agent OS unless explicitly asked.
Validate a checkout with:
scripts/validate.shRun the validator's focused pass/fail checks with:
scripts/validate.sh --self-testGitHub Actions validates pushes and pull requests. After the workflow is pushed, the repository owner must select GitHub Actions as the Pages source if it is not already enabled.
Licensed under the Apache License, Version 2.0. See LICENSE.