What is agent-os?

agent-os is a small set of reusable agent workflows built on the open Agent Skills specification. It currently gives Claude Code and Codex the same working contracts from one shared source.

It answers three questions:

1. What outcome and boundaries govern this work?
2. What may the agent change under the developer's request?
3. What evidence decides whether the result is done?

One managed operating layer supplies policy and skills to both hosts, then checks work against durable repository and evaluation evidence. Select the diagram to open it at full size.

Open standard, tested hosts

The portable core follows the Agent Skills directory model: a SKILL.md file with a name, description, and instructions, plus optional scripts, references, and assets. Skills-compatible agents can read that core format.

Agent OS currently promises tested behavior on two hosts:

• Claude Code uses disable-model-invocation: true to keep manual workflows opt-in.
• Codex uses agents/openai.yaml with policy.allow_implicit_invocation: false for the same rule.
• The npm installer and plugin manifests package those host-specific controls around the shared skill source.

Other skills-compatible agents may be able to load the core instructions, but they are not yet an Agent OS support claim. Wider support requires a real installation and behavior test on that host, not only compatible-looking files. See the current OpenAI skill metadata and Claude Code skill controls for the host extensions Agent OS uses.

Two kinds of skill

Most workflows are manually invoked. They structure coherent planning, bounded shaping, parallel batches, delivery, dispatch, durable lessons, and setup. The automatic check-work workflow is the exception: it activates for supported code-review requests and can also be named explicitly.

Disciplines activate from the situation: diagnose an unknown failure, contain scope drift, verify a completion claim, and treat a developer correction as a misunderstanding signal.

Authority follows the developer

A request to inspect, plan, or review is read-only apart from the requested artifact. A request to implement authorizes in-scope repository changes. Merge, deploy, destructive cleanup, and effects on external systems or people require the request or project policy to include them.

agent-os does not ask for a second permission already carried by the request. It asks when a material product decision remains unresolved.

Compact programs

Skills define objective, working surface, boundaries, ground truth, loop, and exit. They lock the contract while leaving local tactics to the agent. Deterministic operations live in scripts and tests rather than prose.

Everything remains plain Markdown in a Git repository and ships with adapters for both tested hosts.

Next: Getting started.