Three working AI agent governance patterns, free and open source: a decision table, a worked example, and a Microsoft Agent 365 template for each. No sign-up, no trial, no cut-down version, the same content as the paid library, just fewer patterns.

Most AI agent governance content today is either infrastructure (identity, sandboxing, policy enforcement, the plumbing that runs an agent safely) or documentation (a PDF someone has to read and remember). These three patterns are neither. They're executable decision logic for the questions the plumbing doesn't answer:

- Agent Identity Ownership & Sponsor Assignment Your identity platform can enforce zero-trust access. It can't tell you whether the person listed as an agent identity's owner still works here, or was ever the right person. This pattern is that check, as a decision table your systems can run.

- Fleet Budget Ceiling and Circuit Breaker When does agent spend trip a breaker instead of just showing up on next month's bill.

- Agent Inventory Reconciliation Cadence You can't govern an agent identity you don't know exists. This is the cadence that catches the ones nobody registered.

If you're running agent infrastructure with runtime enforcement already in place, these patterns sit above that layer: they're the governance decisions your enforcement tooling needs a rule to check against, not a replacement for it.

- pattern.md— the full pattern: trigger condition, decision logic, provenance, and how it relates to neighbouring patterns.

- decision-table.dmn— the same logic as an executable DMN decision table, tested against every input combination.

- agent365-template.json— a native Microsoft Agent 365 governance template.

- summary.md— a one-page summary of the problem and what the pattern does.

This is 3 patterns out of 92, covering agent identity, cost, and sprawl governance in general. The full library adds 20 packs across lifecycle, pre-deployment, oversight, multi-agent coordination, drift, data governance, change management, and three regulated verticals (employment, worker monitoring, education), all cross-referenced against each other. See outthebox.ai if that's useful to you.

MIT. Use it, fork it, adapt it, ship it. See LICENSE.

This repo is provided as-is. No support commitment, issues and PRs are welcome but may not get a fast response, this is a solo project.