Agent Control Plane Readiness Checklist (2026)

Use this as a pre-flight checklist before you let an agent touch production systems.

1) Identity & Provisioning
- SSO: Support OIDC or SAML with a mainstream IdP (Okta, Microsoft Entra ID, Google Workspace).
- Provisioning: SCIM users/groups so access doesn’t depend on manual invites.
- Service identities: Separate “agent service accounts” from humans; no shared tokens.
- Role model: Define roles (Admin, Builder, Operator, Viewer) with explicit permissions.
- Offboarding: Confirm deprovisioning removes access immediately and revokes tokens.

2) Connector Safety (Tools)
- Scope tokens to the smallest possible API permissions; avoid admin-wide scopes.
- Separate read vs write tools; make read the default.
- Implement “dry run” or “draft” mode for write actions where possible.
- Rate limit tool calls; protect downstream systems from loops.
- Maintain connector versioning and schema checks so vendor API changes don’t silently break behavior.

3) Policy Enforcement
- Central policy engine: allow/deny decisions evaluated at runtime for every tool call.
- Environment awareness: stricter policies in prod than staging.
- Context-based controls: time windows, network posture, project/workspace constraints.
- Approval gates: require explicit approval for destructive or financially material actions.
- Break-glass: an emergency disable switch per workspace, per connector, and globally.

4) Audit & Forensics
- Log every tool call with: actor identity, agent identity, timestamp, inputs, outputs, and policy decision.
- Trace IDs: tie agent actions to downstream system changes (tickets, commits, CRM updates).
- Export: provide log export for security review (common formats are fine; the key is access).
- Data minimization: don’t log secrets; redact where necessary.

5) Evaluation & Release Discipline
- Golden set: maintain a small curated set of real tasks + adversarial prompts.
- CI gating: prompt/policy/tool changes must pass automated evals before deploy.
- Staged rollout: canary new models/prompts to a subset of users/workspaces.
- Regression response: define who rolls back, how, and what triggers a rollback.

6) Cost & Abuse Controls
- Budgets per workspace/team; clear visibility into spend drivers (models, tool calls).
- Throttles: rate limits and max-step limits per run to prevent runaway loops.
- Model routing policy: choose cheaper models for low-risk tasks; restrict premium models by role.

Operator’s Final Check (before production)
- Can you answer: who can run the agent, what it can read, what it can write, and how to stop it?
- Can you produce an audit trail for a single action end-to-end in under 5 minutes?
- Can you prove least privilege on one connector without hand-waving?

If any answer is “no,” you don’t have an agent product yet—you have a demo.