# Plan

## Status

Research complete; report ready for user review.

## Completed work

1. Mapped AI-agent governance products across cloud platforms, identity, MCP/tool gateways, runtime security, lifecycle governance, and observability/evaluation.
2. Mapped generic reusable components for policy, identity, supply-chain trust, secrets, isolation, evidence, compliance, lineage, and approvals.
3. Inspected Microsoft `agent-governance-toolkit` at commit `c63c51e881c442fbc060705f7e211f29993f2c1b` and assessed architecture, enforcement boundary, identity, audit, maturity, and industry position.
4. Compared the market and AGT with the CAF v2 PRFAQ boundary.
5. Produced and independently reviewed `report.md`; corrected the capability taxonomy, AGT semantics, CAF contract scope, and missing build/reuse rows.

## Deliverables

- `report.md`: complete synthesis and recommendation.
- `agent-governance-report.docx`: Word version of the complete report.
- `raw-research/ai-agent-governance-landscape.md`
- `raw-research/generic-governance-components.md`
- `raw-research/microsoft-agent-governance-toolkit-review.md`
- `raw-research/standards-and-frameworks.md`
- `raw-research/caf-v2-build-vs-reuse.md`
- `raw-research/on-prem-self-deployment.md`
- `raw-research/policy-language-landscape.md`

## Working recommendation

- Build CAF's product-defining semantic layer.
- Reuse host policy, identity, credentials, scheduling, isolation, telemetry, retention, and incident response.
- Keep Pi pinned and unforked for the MVP.
- Pilot a fixed AGT/ACS version behind a replaceable policy-provider interface and compare it with a minimal OPA/Cedar alternative.
- Use second-runtime admission as the next decisive architecture proof.
- Treat full on-prem self-deployment as a first-class selection criterion; distinguish a customer-managed control plane from local connectors, private endpoints, or hybrid data planes.

## Next checkpoint

User review of the report and decision on whether to authorize the bounded technical spike.

The Markdown-to-DOCX converter is preserved at `scripts/md_to_docx.py` for repeatable regeneration.