agent2

# Agent2

> Agent2 turns domain experts into production AI agents. It clones how
> professionals work: their workspace, books, tools, memory, clarification
> loops, approval flow, and typed work product.

Agent2 is built on PydanticAI and FastAPI. Open this repo in Claude Code, Codex,
Cursor, Gemini CLI, or another coding agent and use `/brain-clone` for serious
domain agents.

v0.3 adds a first-class CLI onboarding path:

- `agent2 setup`: write `.env` and `agent2.yaml`, select model/profile/telemetry.
- `agent2 onboard`: run the Brain Clone harness and generate an Agent2 agent.
- `agent2 doctor`: validate Docker, uv, config, ports, compose, health.
- `agent2 list`, `agent2 run <agent>`, `agent2 serve <agent>`, and
  `agent2 publish-check`: operate local agents.

Model resolution is explicit runtime argument > agent `config.yaml` >
`agent2.yaml` > env fallback. Normal agents should leave `model: ""`.

The most important distinction: prompts teach the expert's Sachbearbeiter Chain-of-Thought; knowledge
books contain what the expert knows. Do not replace the knowledge layer with
hardcoded lookup tables.

## Primary Files

- [AGENTS.md](./AGENTS.md): root instructions for coding agents.
- [CLAUDE.md](./CLAUDE.md): Claude Code entry point that imports AGENTS.md.
- [llms-full.txt](./llms-full.txt): expanded framework context.
- [shared/runtime.py](./shared/runtime.py): `create_agent()` and model setup.
- [shared/api.py](./shared/api.py): `create_app()`, task API, hooks, resume,
  per-run toolsets, and approval endpoint.
- [agent2_cli/](./agent2_cli): Typer/Rich/Textual-ready onboarding CLI,
  setup wizard, doctor, deterministic Brain Clone generator.
- [agent2.yaml](./agent2.yaml): global framework configuration.
- [knowledge/collections.yaml](./knowledge/collections.yaml): knowledge
  collection catalog.

## Canonical Patterns

- [Brain Clone Pattern](./docs/brain-clone-pattern.md): how to build full domain
  expert agents.
- [Sachbearbeiter Pattern](./docs/reference-agents/sachbearbeiter-pattern.md):
  production-proven reference architecture.
- [Procurement Compliance Officer](./agents/procurement-compliance-officer):
  full in-repo flagship example using Knowledge MCP, scoped per-run toolsets,
  three outcomes, memory, approval, resume, `after_run`, mock mode, and evals.

## Demos

- [approval-demo](./agents/approval-demo): pending actions.
- [resume-demo](./agents/resume-demo): message history.
- [provider-policy-demo](./agents/provider-policy-demo): provider affinity.
- [scoped-tools-demo](./agents/scoped-tools-demo): tool/collection scoping.
- [example-agent](./agents/example-agent), [support-ticket](./agents/support-ticket),
  [code-review](./agents/code-review), [invoice](./agents/invoice), and
  [rag-test](./agents/rag-test): small reference demos.

## Docs

- [README](./README.md)
- [Architecture](./docs/architecture.md)
- [CLI Onboarding](./docs/cli-onboarding.md)
- [Creating Agents](./docs/creating-agents.md)
- [Capabilities](./docs/capabilities.md)
- [Knowledge Management](./docs/knowledge-management.md)
- [Approvals](./docs/approvals.md)
- [Resume](./docs/resume-conversations.md)
- [Provider Policy](./docs/provider-policy.md)