AI coding agent usage policy template, with each rule enforced

the short answer

Yes, below. Most AI usage policies are documents nobody can enforce. This template pairs every clause with the mechanism that makes it true: managed settings, a hook-level policy, credentials or a review step, and it assumes several approved agents rather than one. Adopt the clauses that fit, delete the rest, and only publish a clause once its control is deployed, at least in observe mode.

Format: Clause, enforcement mechanism, evidence
Sections: Approved agents, accounts, data, actions, credentials, review, monitoring
Rule: No clause without a control
Multi-agent: One list of approved agents; one policy layer across them
Owner: Platform engineering with security

Write It for Several Agents, Not One

Most engineering teams now run more than one agent: Claude Code in the terminal, Cursor in the editor, Codex or Copilot for some teams. A usage policy that names one tool is out of date the week it ships. Name the approved agents and plan tiers in one list, apply each vendor’s own admin settings to pin accounts and modes, and put the action rules in a policy layer that covers all of them.

The Template

Clause	Enforced by	Evidence
1. Only approved agents and plans are used on company code: [list, e.g. Claude Code (Team), Cursor (Enterprise), Codex (Business)].	Inventory script; policy client installed on every machine	Inventory report; coverage under Admin → enforcement
2. Agents sign in only to company accounts.	Claude Code `forceLoginMethod` and `forceLoginOrgUUID`; Codex `allowed_chatgpt_workspaces` (requirements.toml or MDM); Cursor `AllowedTeamId`	Managed settings and MDM profiles
3. Agents do not read secret files or print environment secrets.	`block-env-files`, `protect-env-vars` (on by default)	Deny decisions
4. Agents do not run destructive infrastructure commands.	Custom policies (see the Terraform, Route 53 and database guides)	Deny decisions
5. Agents do not push directly to protected branches or force-push.	`block-push-master` (set `protectedBranches`), `block-force-push`	Deny decisions; branch protection
6. Agent work in the cloud uses the approved agent profile.	Approved-profile custom policy; IAM permission set	Policy version; IAM config
7. Infrastructure changes from agents go through pull requests.	Deny local applies with a reason that points to a PR	Deny decisions; PR history
8. Agent activity is recorded for every session.	Policy client connected to the fleet console	Session traces
9. Bypass and auto-approve modes are disabled.	Claude Code managed settings: `permissions.disableBypassPermissionsMode: "disable"` and `disableAutoMode: "disable"`; equivalents per agent	Managed settings files
10. Exceptions are requested, approved and time-limited.	Per-machine deployment (persona assignment in managed deployments), rolled back with `fp fleet rollback` when the exception ends	Deployment history

How to Adopt It

Delete clauses you will not enforce. An unenforced clause trains people to ignore the document.
For each remaining clause, deploy the control in observe mode (a Cloud deployment) and read a week of decisions.
Publish the policy and switch controls to enforce in the same week.
Review quarterly against new agents and new incidents.

Two Commands Cover Clauses 3 and 5

block-push-master protects main and master unless you set protectedBranches under policyParams in policies-config.json; add your release branches there.

failproofai policies add FailproofAI/policies   # env files, env vars, protected branches, sudo, curl | sh on by default
failproofai policies add block-force-push       # off by default

FAQ

Do developers need to sign it?

Many companies attach it to the existing acceptable use policy acknowledgement. The controls matter more than the signature.

How specific should the approved tools list be?

Name the tools and the plan tier, for example “Claude Code on the company Team plan”. Generic wording invites personal accounts.

What about non-engineers using agents?

The same template works; trim the infrastructure clauses and keep the data and account clauses.

Sources

Checked against the sources below on October 21, 2026. Harness settings, hook capabilities and cloud features change between versions.