What should an agency hand a client who asks who checked the agent?

Hand them evidence they can read without you in the room. That means a written list of what the agent may and may not do, the points where it must stop for a human, spend limits, and a file set whose integrity anyone can verify. If the client needs more than your word, add a review by someone who did not build the agent.

Key takeaways

• The question arrives at sign-off, so the answer has to exist before the demo.
• Write the forbidden actions first. Capabilities are the easy part.
• A pack your client can verify beats a slide that says "secure by design."
• Your own review is useful, but it is not independent. Say so plainly.
• Paper does not enforce anything. The runtime has to check every action against it.

Why does this question land on the agency?

Picture the last meeting before launch. The demo went well. The agent answered tickets, drafted replies and booked a call on its own. Then the client's operations lead asks one quiet question: who checked what this thing is allowed to do?

Most agencies answer with a person. "Our lead engineer tested it." That is true, and it is also the problem. The person who built the agent is vouching for the agent. The client has no file to read, no list to compare against and nothing to hand their own boss.

The agency is where this question lands because the agency is the last party that can still change the build. After handover, the client owns the risk, and they know it.

What does a good answer look like on paper?

A good answer is a short stack of documents the client can open, read and check later. In ClawMaven, that stack is a governance pack: a ZIP of 50+ files generated in your browser. The parts that matter most to a client are these.

• Policy. Approved actions, forbidden actions, budget caps and who may operate the agent.
• Approval gates. The exact moments the agent must stop and ask a human.
• Identity. The agent's persona, boundaries and behavioural limits.
• Trust manifest. A SHA-256 hash of every file, so anyone can prove later that nothing was edited.
• Verify scripts. For macOS, Linux and Windows, so the client can recompute every hash themselves.
• Compliance summary. Mapped to the EU AI Act, NIST AI RMF, ISO/IEC 42001, SOC 2 and GDPR.
• Deployment guide. How the pack attaches to the runtime you used.

The wizard has an Agency edition for this kind of work. You build one pack per client agent, and each pack stands on its own.

What should you write first?

Write the forbidden-action list before anything else. Every agency can describe what its agent does. Fewer can describe what it must never do.

Start with the actions a client would be embarrassed to explain. Sending email to people outside an approved list. Spending above a set amount. Reading files outside its folder. Reusing credentials it finds. Changing its own instructions.

Then attach an approval gate to anything that leaves the client's network, moves money or touches credentials. A gate is a sentence a client can understand: "Before any refund above $50, the agent stops and asks the account owner." If you cannot write the gate in one plain sentence, the client will not trust it.

Is your own review enough?

Sometimes it is. A small client with one low-risk agent may be happy with your pack and a walkthrough.

Larger clients often are not, because a self-check has a built-in conflict. You are grading your own work. Procurement teams, insurers and compliance leads know that, even when they do not say it.

Be honest about which kind of review you are offering. "We wrote this policy and tested the agent against it" is a fine statement. It is a different statement from "an outside party reviewed what we shipped and put their name on the finding." Do not let the first one pass for the second.

Does a pack stop the agent from misbehaving?

No, and no honest vendor should claim it does. ClawMaven is not a runtime firewall. A pack defines the rules before production, in files your runtime reads. The runtime then has to check each action against that policy.

That is why the deployment guide matters. When you hand over the pack, show the client where the runtime enforces it. If nothing enforces it, tell them that too. A policy no system reads is a promise, not a control.

It also helps to be plain about the law. No law requires this kind of outside evidence today. Under the EU AI Omnibus, the stand-alone high-risk obligations now apply from 2 December 2027, while Article 50 transparency obligations have applied since 2 August 2026. The practical driver for most agencies is simpler. Clients are starting to ask before they sign.

What should you do this week?

• List every agent you have shipped that can send messages, spend money, write files or call an outside API.
• For each one, write the forbidden-action list and the approval gates in plain sentences.
• Generate a pack and run the verify script once yourself, so you know what the client will see.
• Decide, per client, whether your own review is enough or whether they need an outside one.

You can build a pack free at clawmaven.com, with no account and no credit card. If you run several client agents, Team at $99 a month adds team sharing, config versioning and branded runbooks. If a client needs an outside review, the Governance Audit is fixed scope, from $2,500, delivered in 5 business days: a posture score, a compliance summary, a deployable pack, a trust manifest and a leadership briefing. If they want someone checking the agent against its pack every month after launch, that is ClawMaven Observer at $497 a month.

Clients do not need you to promise the agent is safe. They need something they can check without you.

Frequently asked questions

What should an agency give a client before an AI agent goes live?

A written policy with approved and forbidden actions, approval gates for risky steps, budget caps, and a file set the client can verify. A ClawMaven governance pack covers these in one ZIP, with a SHA-256 trust manifest and verify scripts for macOS, Linux and Windows.

Is an agency's own review of its agent enough?

It can be for a small client with a low-risk agent. It is not independent, though, because the builder is checking its own work. When a client's procurement, insurer or compliance lead needs more, an outside review by someone who did not build the agent answers the question better.

Does a governance pack enforce the rules by itself?

No. ClawMaven is not a runtime firewall. The pack defines permitted and forbidden actions, approval gates and limits before production, and your runtime has to check each action against that policy. The deployment guide shows how the pack attaches to the runtime you used.

How can a client check that the governance files were not changed?

Every pack includes a trust manifest with a SHA-256 hash of each file, plus verify scripts. The client runs the script, it recomputes every hash, and any edit after handover shows up as a mismatch.

What does the ClawMaven Governance Audit include?

It is fixed scope, from $2,500, delivered in 5 business days after a 30-minute discovery call. It includes a governance posture score from 0 to 100, a 10-section compliance summary, a deployable governance pack, a SHA-256 trust manifest and an executive leadership briefing.