Document-defined agents

In Airic, agent behavior is defined by documents — roles, processes, precedents, and tool policies stored as Markdown files in your workspace.

Why documents?

Modern agents can code, research, and automate. But serious work depends on methodology: how to frame problems, make decisions, review results, and improve over time. That knowledge should be:

• Editable — change a role spec, not a buried system prompt
• Versionable — track methodology changes in git
• Inspectable — anyone on the team can read what the agent will do
• Composable — packs combine roles, document types, processes, and precedents

Spec types

Spec	Purpose	Example path
Roles	Who the agent is and how it decides	specs/roles/reviewer.md
Document types	Schemas for task and output documents	specs/document-types/task.md
Processes	Step-by-step execution flows	specs/processes/discovery-loop.md
Precedents	Captured judgment for future runs	specs/precedents/jtbd-framing.md

Extension packs

A pack packages a full methodology:

packs/<your-method>/
├─ roles/
├─ document-types/
├─ processes/
├─ precedents/
└─ tool-policies/

Packsmith helps you build packs from source material. See Packsmith for the extraction pipeline.

Executable documents

Airic turns clarified intent into task documents agents act on — durable artifacts, not ephemeral chat messages:

task.md → run()

The agent loads the relevant pack, selects a role and process, and produces output documents you can review before committing.