Policy starter

AI Action Boundary Map

A practical allow / approve / block starting point for AI-assisted software delivery actions. Use it to keep normal coding fast while gating paths with blast radius.

Last updated: May 13, 2026

See it as a graph

The map below is a starter policy, not a universal rule. The right boundary depends on credential scope, target system, reversibility, owner, approval model, and evidence coverage.

Starter action boundary

Allow

  • Code edits in normal repo scope
  • Tests, linting, and docs
  • Local refactors
  • Read-only repo analysis
  • Non-sensitive generated comments

Approve

  • Package installs or dependency changes
  • Workflow or CI/CD config edits
  • MCP calls to internal systems
  • Deploy or release triggers
  • Cloud CLI or infrastructure commands
  • Secret-adjacent reads

Block

  • Disabling required checks
  • Unmanaged secret reads
  • Broad PAT or static key use without purpose
  • Destructive production actions
  • Workflow changes that bypass approval

Evidence to keep at the boundary

Owner and purpose

Who asked for the action, which agent or automation acted, and what business or engineering purpose was declared.

Credential and scope

Which token, service account, OAuth grant, CI secret, cloud role, or inherited identity was involved.

Approval reason

Who approved the action, what policy applied, and why it was allowed, approved, or blocked.

Outcome

What changed, which target was affected, what validation ran, and where logs or proof live afterward.

Copyable map

See how this boundary appears in a graph.

The lab shows where allow, approve, or block decisions appear across agent, repo, workflow, credential, action, target, and proof.

Open the lab