Skip to content
be-human.ai
← Back to the journal

Harness fundamentals

Human Approval Gates for AI Agents: Where to Place Them in a Harness

Learn where to place human approval gates in an AI agent harness so irreversible actions wait for a named person while low-risk steps run freely.

Be Human typographic cover with the title: Human Approval Gates for AI Agents: Where to Place Them in a Harness
Conceptual illustration created with AI · Be Human

A human approval gate for AI agents is a checkpoint in the agent harness that pauses one specific action until a named person approves or rejects it. Place gates on actions that cannot be undone, that spend money, that change customer-facing records, or that expose personal data, and let read-only and reversible steps run without waiting. The rest of this guide shows how to find those points, where to put them in the execution path, and how to keep the pauses useful rather than ignored.

Key takeaways

  • Gate actions by their consequences, not by the name of the tool that performs them.
  • Put the gate in the harness, outside the model, so a prompt cannot talk its way past it.
  • Each gate needs one accountable approver, a backup, a deadline, and a default outcome of no action.
  • Approve the exact payload the agent proposed, and treat any change as a new request.
  • Log every request, decision and reason so you can audit the gate later.

What an approval gate is inside an agent harness

A harness is the software around the model that decides what the agent can see, which tools it can call, what it remembers, and when it stops. The model proposes the next step. The harness decides whether that step runs. An approval gate is a rule in that second layer: before a tool call with a given impact executes, the harness checks whether a person has approved this specific call.

This distinction matters. A line in the system prompt that tells the agent to ask before sending refunds is a request to the model, and models can misread, forget, or be steered by content they read during a task. A gate enforced by the harness does not depend on the model's cooperation. The model can propose a refund all it likes, but the request sits in a queue until someone with authority says yes. The OWASP Top 10 for LLM Applications describes excessive agency, where an agent holds more permission or autonomy than its task needs. The practical response is to limit what the agent can do, and not to rely on instructions alone to limit it.

Sort every action by reversibility and impact

Before you place any gate, sort the actions the agent can take. Most of the work happens in this sorting, and it is easier to do with a plain list than with a long policy document.

  • Read-only lookups, such as checking an order status or reading a calendar, run freely.
  • Drafts that stay inside the team's workspace run freely, because nothing leaves the business.
  • Reversible internal writes, such as adding a note or tagging a record, run freely with full logging.
  • Messages to customers, vendors or outside parties need approval before they are sent.
  • Payments, refunds, deletions, permission changes, publishing and bulk updates need approval from a named owner.
  • Any action that moves personal data into a new purpose or a new recipient needs review before it runs.

Two questions settle most cases: can this be undone within the normal business process, and would a wrong result reach someone outside the company? If either answer is no, gate it. For more on which controls to set first, see AI agent guardrails for business. Keep the gated list short enough that approvers can actually read every request.

Where to place the gate in the execution path

The gate belongs between the model's proposed tool call and the tool executor. The following sequence works for a first implementation.

1. List every tool the agent can call, including the ones inherited from integrations you had forgotten about. 2. Tag each tool, and each argument pattern that matters, with an impact level taken from the sorting above. 3. Add a policy check in the harness that runs after the model proposes a call and before the executor touches an outside system. 4. When a call is gated, store the full request with its arguments, the context the agent used, and the reason it was flagged. Pause that branch of the work instead of letting the loop continue. 5. Notify the approver through the channel your team already watches, with enough context to decide without opening three other tools. 6. On approval, execute exactly the stored payload. On rejection, return the reason to the agent so it can revise the request or stop. 7. Record the outcome, the approver and the time, then resume or close the task.

Teams often gate at the wrong layer. Gating only the final message misses an agent that already changed a record. Gating only the model's text misses tools that run on their own. The check should sit where the side effect happens. Checks that run before and after each step are covered in AI agent hooks.

Name one approver, a backup and an expiry

An approval that anyone can give is effectively no approval. Each gate should name one accountable person, usually the owner of the process the action affects, and one backup for absences. The harness should route requests to that person, not to a shared inbox where each message waits for someone else to act. Authorization guidance from OWASP stresses checking permission at the point of action, so the harness should verify the approver's rights when the decision arrives instead of assuming them from the original request.

Every request also needs an expiry. If nobody responds, the default outcome must be no action, and the affected party should hear that the step is pending or closed. A request that sits for days and then executes because nobody answered is a failure, not a convenience.

Illustrative scenario: an online shop's support agent drafts a refund for a delayed order. The refund tool is gated. The operations lead receives the order number, the amount, the customer's message and the reason the agent flagged the case. If the lead does not decide by the end of the business day, the request expires, the refund does not run, and the customer receives a status message that the case is under review. The next morning, the lead decides with full context instead of approving something stale.

Bind each approval to the exact action

The most common gap in approval flows sits between what was approved and what was executed. A person reviews a proposed change, the agent revises its arguments while the request waits, and the executor runs the newer version. To close this gap, attach the approval to a fingerprint of the exact payload. If the payload changes, the harness creates a new request.

Idempotency covers the second risk. If an approval response arrives twice, or the executor retries after a timeout, the action must not run twice. Payment providers document this pattern: the client sends a unique key with a request, so a repeated call returns the original result instead of creating a second charge. Apply the same idea to every gated write. Use a stable request identifier, and make the executor check whether that identifier has already been processed. Verify the source of approval callbacks as well, so a forged or replayed message cannot release a gated action.

Keep the gates from becoming rubber stamps

A gate that is approved every time for months teaches people to click without reading. Review the gate log on a regular schedule and look at three things: how many requests each gate receives, how often approvers reject them, and whether the agent ever proposed an action that went through ungated because its impact tag was wrong.

If a gate is never rejected and is always approved within minutes, consider whether it should become a notification rather than a pause. If a gate is rejected often, the agent probably needs better inputs or narrower instructions, not a permanent human bottleneck. Changes to gate rules are harness changes too. Record them, review them and watch for drift, because a permission widened quietly in one release is the most common way an approval point disappears without anyone deciding to remove it. The pattern is covered in harness drift and unreviewed changes.

How Be Human handles approval gates in managed operations

Be Human designs and operates agents with these gates defined in advance. A first engagement defines one useful role, the memory and tools it needs, and the actions that require a named approver, each with a backup and an expiry rule. Managed operations then review permissions and gate logs, watch for requests that never resolve, test changes before they reach live systems, and respond when a gate fails or is bypassed. Gates reduce risk but do not remove it, so each pause is made visible, owned and reviewable.

If you want to find which of your processes is ready for an agent and where its first gate should sit, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Which actions should always need human approval for an AI agent? Actions that cannot be undone within your normal process should be gated first, including payments, refunds, deletions, permission changes, and anything published or sent to people outside the company. Add any action that moves personal data into a new purpose or a new recipient. Start with the smallest set you can defend, and widen it only after the gate log shows where the real risk sits.

Can a prompt instruction replace an approval gate? No. A prompt is guidance the model may follow or misapply, while a gate is a rule the harness enforces before the tool runs. Use instructions to shape how the agent asks for approval, such as including the reason and the affected record, but keep enforcement in the harness, outside the model's reasoning.

What happens if the approver does not respond? The request should expire at a deadline you set in advance, and the default outcome must be no action. The affected party should receive a plain status message, and the case should return to a human queue with its full context. An unanswered request should never execute simply because time passed.

How many approval gates should a first agent have? Usually fewer than teams expect. A first agent should have a small set of gates on high-consequence actions, with everything else running freely under logging. Too many gates produce fatigue and quick approvals that defeat the purpose. Add gates after reviewing the log, and remove them only when a pattern has been reviewed and proven consistently safe.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →