Agent OS is an operating system for coordinating AI agents as a team.
It defines the durable parts of an agent organization:
- core: operating principles and boundaries
- workflows: loops for action, review, and improvement
- agents: installable overseer, council, and expert agents
- skills: reusable capability packages
- adapters: runtime/provider-specific implementation notes
- examples: sanitized deployment patterns
The goal is not to publish a prompt pack. The goal is to make agents safer, more useful, more reviewable, and easier to scale without creating chaos.
Every durable agent has an improvement workflow: it acts, reviews what happened, and improves its own behavior or reports system-level improvements to the overseer.
Pre-release. The repository is being audited for correctness, installability, and public launch readiness.
Agent OS uses a front-door overseer agent, independent council agents, and scoped expert agents.
The overseer owns:
- user-facing continuity
- triage and routing
- synthesis across agents
- escalation and safety gates
- durable memory hygiene
Expert agents own bounded work:
- research
- product analysis
- engineering
- growth
- finance
- review
- operations
core/— principles and system modelworkflows/— binding workflowsagents/— installable agent charters and workspace filesskills/— reusable capabilities agents can invokeadapters/— provider/runtime profilesexamples/— sanitized deployment sketchestemplates/— reusable document templates
Generic core belongs here. Private operating state does not.
Do not publish:
- personal facts
- customer or venture-specific context
- secrets or secret-store names
- real host paths
- private agent memory
- live operational logs
- account IDs, emails, tokens, credentials, or API keys
Use tokens and placeholders instead.