Agent Run Config

Portable, governed AI agents.

AgentRC is an open specification for declaring, packaging, securing, and sharing AI agents. It defines the contract an agent declares; compatible runners decide how to execute it.

```dockerfile # syntax=agentrc.agentfile/v0.1 AGENT code-reviewer CMD claude --print TOOL utci:file_read TOOL utci:shell MOUNT /workspace rw SECRET github_token env:GITHUB_TOKEN AUDIT all POLICY permit( principal == AgentRC::Agent::"code-reviewer", action == AgentRC::Action::"tool.invoke", resource == AgentRC::Tool::"file_read" ); END ```

Not a runtime

AgentRC does not implement containers, microVMs, cloud sandboxes, or model loops. It declares what must be true before a runner executes an agent.

Security by declaration

Tools, mounts, network egress, credentials, rate limits, and audit requirements are declared in a reviewable file and pinned into a package.

Registry-native

AgentRC packages are designed for OCI-compatible registries, so agents, bases, tools, policies, and skills can be shared like container images.

The separation AgentRC creates

Concern Defined by Implemented by
Agent identity and entrypoint Agentfile agent author
Tools, skills, functions, MCP servers Agentfile + lockfile package builder / runner
Security boundaries policy and declarations compatible runner
Packaging and sharing AgentRC package profile OCI registry
Execution substrate runner profile Docker, gVisor, Firecracker, cloud runners, local runners
Multi-agent workflow future workflow profile workflow engines
Core claim: The Agentfile declares one agent. The lockfile pins dependencies. The package makes it portable. The policy makes boundaries reviewable. The registry makes it shareable. Compatible runners execute it.

Current draft

Working Draft 0.1 Agentfile OCI package Cedar profile Runner conformance Workflow draft

The project is ready to publish as a standards-style repository: specification first, reference tooling second.