Mel – the agentic terminal
The agentic team for coding.
A one-page AI coding agent policy for your engineering team (template)
November 22, 2026 · 6 min read

- A one-page policy says what an AI coding agent may do alone and what needs a human yes.
- Ten rules cover scope, keys, secrets, approvals, branches, review, ownership, data, incidents and review dates.
- Roll it out in a two-week pilot with two or three people, then fix the rules that did not work.
- Mel – the agentic terminal has configurable approvals, so each developer can set them to match the policy.
A one-page policy tells your developers what an AI coding agent may do alone and what needs a human yes; Mel – the agentic terminal (openmel.dev) lets each team member set approvals to match it. Copy the ten rules below, fill in the blanks, get one sign-off, and try it for two weeks before you make it final.
Why does a team need an AI coding agent policy?
Because without one, every developer makes up their own rules. One person lets the agent run anything. Another approves every single line and gives up after a day. A third pastes customer data into a prompt without thinking about it.
None of these people are careless. They just have no shared answer to three questions:
- What may the agent do without asking?
- What must a person approve, every time?
- Who is responsible for what the agent changed?
A written page answers all three. It also gives new hires and reviewers something to point at.
What goes in a one-page AI coding agent policy?
Here is a template. Copy it into your wiki or your repository, replace the blanks, and delete any line that does not fit your team.
AI CODING AGENT POLICY (team: ____, version 1, date: ____)
1. Scope: agents may work in these repositories: ____. Not in: ____.
2. Models and keys: use only these model providers: ____. Keys come from ____.
3. Secrets: no production secrets or customer credentials in any folder
an agent can read. Use test keys and sample data.
4. Approvals: start with approvals on for commands and edits. Loosen them
only for routine tasks listed here: ____.
5. Always a human yes: pushes, deploys, deleting data, changing CI,
adding dependencies. Check your tool's settings: never assume it asks.
6. Branches: agent work happens on a branch, never directly on main.
7. Review: before a commit, the developer reads the full diff and runs
the tests. Agent work then goes through normal code review.
8. Ownership: the person who started the agent owns the commit.
9. Data: no customer personal data in prompts or pasted logs.
10. Incidents: report any surprising agent action in ____ within a day.
We review this page after 2 weeks, then on the first Monday of each month.Rule 5 is the one teams get wrong most often. Writing "deploys need a human yes" does not make a tool ask. Each developer has to open their own tool's approval settings and confirm it, and your git host, CI and review process are the second line.
Who should sign off the policy?
Keep it small. One engineering lead owns the page and signs it. If your company has a security or IT person, ask them to read rules 2, 3 and 9, since those are about keys and data. If you work with customer data under a contract, ask whoever owns that contract to read rule 9.
Then share it with the whole team and ask for objections within a week. A rule that people quietly ignore is worse than no rule, so the objections are useful.
How do you roll the policy out in a two-week pilot?
Do not roll it out to everyone at once. Pick two or three developers and one repository.
Week 1. Approvals on for commands and edits. One kind of task only, for example bug fixes with a failing test. Everyone keeps a short note of each time the agent surprised them or each time a rule got in the way.
Week 2. Add a second kind of task. Let each person loosen approvals for the routine steps they listed in rule 4, and only those. Keep rule 5 as it is.
End of week 2. Read the notes together. Change the rules that slowed people down for no reason, keep the ones that caught a real problem, and publish version 2 to the whole team.
How do you check that the policy is followed?
A policy you cannot check is a wish. Each rule should have a simple check.
| Rule | Why it matters | How to check |
|---|---|---|
| Scope | Keeps agents out of code they should not touch | List the repos each person used the agent in during the pilot |
| Models and keys | You know which providers see your code and who pays | Compare keys in use with the approved list |
| Secrets | An agent can send what it reads to the model | Search the repos for secrets before the pilot starts |
| Approvals | People stay in control of commands and edits | Each developer shows their approval settings once |
| Always a human yes | The costly mistakes happen here | Try each step on a test repo and see if the tool asks |
| Branches | Mistakes stay off main | Check that agent work arrives as branches |
| Review | The commit is where agent work becomes your work | Ask in code review: "did you run the tests yourself?" |
| Ownership | Someone answers for every change | The commit author is the person who ran the agent |
| Data | Customer data stays out of prompts | Spot-check shared prompts and pasted logs |
| Incidents | You learn from surprises before they repeat | Count reports at each policy review |
The "Always a human yes" check is the important one. Test it yourself on a throwaway repository instead of trusting a settings screen or a product page, including ours.
Where does Mel fit?
Mel – the agentic terminal is the agentic team for coding: a multi-agent, multi-model agentic terminal where a team of agents works on files, commands, code and diff review. Configurable approvals keep you in control of commands and edits, so each developer can set them to match rule 4 of your policy.
For rule 2, Mel works with Claude, GPT, Gemini, DeepSeek and Qwen, so you can list US and Chinese models in one tool and pick from the providers your policy allows. Mel is free with your own API key on Mac, Windows and Linux, plus a VS Code extension.
For a team pilot, Mel Business is $50 a seat with two seats minimum, shared credits and one invoice, which keeps rule 2's "who pays" answer simple.
Mel does not replace the rest of the page. Branches, review, ownership and the incident note are habits your team keeps.
Frequently asked questions
Do we need a policy if we already do code review? Yes. Code review sees the result. The policy covers what happens before it: which commands the agent ran, which keys it used and what data went into the prompt.
Should approvals be the same for everyone? Start the same for everyone. After the pilot, let experienced developers loosen approvals for the routine steps you listed, and keep the "always a human yes" list the same for the whole team.
Can developers use their own API keys on company code? That is your call, and rule 2 is where you write it down. If you allow it, say which providers are allowed and who pays.
Does Mel enforce our policy for us? No tool enforces a whole policy. In Mel, configurable approvals keep you in control of commands and edits; your branches, CI and code review cover the rest.
How long should the policy be? One page. If it grows past that, move the details to a separate how-to and keep the ten rules on top.
Try Mel
Download Mel for free and try your draft policy on a test repository with approvals on. If you want to walk through the policy with your team on your own code, book a demo.
Mel – the agentic terminal · openmel.dev. The agentic team for coding.

Lalla is a hands-on tech founder, agentic infrastructure architect and the vision behind Mel.
Follow on LinkedIn