How can we set useful human approval gates for an AI agent?

An agent that prepares a message presents a different decision from one that sends it. Likewise, suggesting a purchase differs from committing funds. We recommend placing human review at these transitions, where preparation becomes an action with consequences beyond the immediate task.

For solo founders, small teams, nonprofits and creators, a useful starting point is to distinguish work we are comfortable delegating from decisions we want to retain. Approval gates turn that distinction into a requirement we can document for runtime implementation.

When should we ask for human approval?

We recommend considering approval when an action spends money, changes access, shares sensitive information or makes an external commitment. These are planning considerations, not claims about what a particular agent or platform supports.

We can begin with three questions: What could this action affect? How difficult would it be to reverse? Who has the authority to accept the consequences? The answers help us decide whether the agent should proceed within a defined limit, prepare a draft or pause for review.

For example, we might permit drafting an outreach message but require approval before sending it. We might also allow preparing a purchase recommendation while reserving the spending decision for a person. The useful distinction is between preparing work and authorizing its effects.

What should an approval request explain?

We suggest documenting the proposed action, the affected resource, the reason for the request and any relevant spending or access limit. A reviewer needs enough context to understand the decision rather than simply acknowledge a notification.

We should also specify what happens while a request is pending. A proposed policy might require the agent to pause the affected action until an authorized person responds. We need the runtime to implement that behavior; writing it into a package does not make it happen.

It is worth defining who may approve and what changes require another review. If the proposed recipient, tool or cost changes after approval, we can treat that as a reason to reconsider the authorization rather than assume the earlier decision still applies.

How can we document these choices with Agent Bob?

Agent Bob is a governance package builder by Master 22 Solutions, powered by ClawMaven. We can use its 7-step wizard to work through purpose, personality, tools, autonomy, budget, memory and approvals. The builder is free, with no credit card required.

We recommend considering approvals alongside the other six areas. Tool access affects what an agent could do, autonomy defines the intended scope of delegation, and budget adds a spending boundary. Memory decisions can also prompt questions about what information should be retained.

Bob Free costs $0 and creates unsigned draft packages. Bob Pack costs $29 one-time per agent and includes a signed governance package. Expert review costs $99 one-time, includes Bob Pack and a written review of gates, tools and limits, with written fixes within 2 business days and no calls needed.

What still needs to happen in the runtime?

Agent Bob does not host or run agents, provide live monitoring or enforce policy files by itself. A runtime must implement the policy and approval gates. SHA-256 integrity and HMAC signing do not prove runtime compliance. ClawMaven provides separate runtime governance.

We can read more in the AI agent governance guide. When considering frameworks such as NIST AI RMF, ISO/IEC 42001 and the EU AI Act, we should distinguish mapping from compliance or certification. Mapping to a framework is neither.