Problem-Solution Page

Your AI Agent Can Act. You Do Not Know Where It Should Stop.

An agent that can send emails, update CRM records, issue refunds, book meetings, or delete files needs defined stop points. Not because the AI is untrustworthy, but because every system that can act on your behalf needs to know which actions need human approval and which can run without it.

4
Control Layers
3
Action Risk Levels
6
Warning Signals
1
Architecture Path

The Core Problem

3 Action Risk Levels Every Agent Needs Defined

Level 1: Run freely

Reversible, low-impact

Drafting content, reading data, generating summaries, classifying records. These can run without approval. If wrong, a human can correct the output with no cost.

Level 2: Suggest and confirm

Potentially visible or moderately impactful

Sending a customer email, updating a CRM field, scheduling a meeting. The agent prepares the action. A human reviews and confirms. One approval step before it goes out.

Level 3: Human decides

Irreversible or high-value

Issuing refunds, cancelling orders, deleting records, modifying contracts, sending bulk messages. The agent prepares a recommendation. A named person approves before anything executes.

4 Control Layers

What Every Production Agent Needs Before It Runs

01

Action scope definition

Write a list of every action the agent can take. For each one, assign a risk level: run freely, suggest and confirm, or human decides. If you cannot list the actions, the scope is undefined. An agent with undefined scope can take actions that were never intended.

Good sign

You can list 5 to 15 specific actions and assign a risk level to each in under 20 minutes.

Bad sign

The answer is "it depends on the task" with no defined list. That is not a scope. That is an open-ended tool.

02

Human approval gate design

For every Level 2 and Level 3 action, define who approves it, how they receive the approval request, how long they have to respond, and what happens if they do not respond. A timeout that defaults to "proceed" on no response is not a gate. It is a rubber stamp.

Good sign

Each Level 2 and 3 action has a named approver, a delivery channel, a timeout, and a defined fallback on timeout.

Bad sign

The timeout defaults to "proceed" or the approval step was skipped to save development time.

03

Audit log for every action

Every action the agent takes should create a log entry: what it did, what data it used, what the outcome was, who approved it if approval was needed, and when. Without logs, you cannot tell customers what happened, investigate a complaint, or prove the agent behaved correctly.

Good sign

You can pull the last 100 agent actions in under 2 minutes and see what each one did and why.

Bad sign

No structured logs. Or logs that show the output but not what data the agent used to produce it.

04

Rollback path for high-risk actions

For every Level 3 action, define how to undo it. A refund issued incorrectly can be reversed with Stripe. A customer email sent to 500 people cannot be unsent. Know which actions are reversible and design the rollback procedure before the action runs in production, not after the first mistake.

Good sign

Each Level 3 action has a documented rollback procedure and it has been tested at least once.

Bad sign

The rollback assumption is "we will handle it if it happens." That is not a rollback path.

Diagnosis

Warning Signs Your Agent Needs Control Work

What you seeWhat it usually meansWhat to do next
Agent sent an email or message you did not expectNo approval gate on outbound actions, or scope not defined explicitlyAdd approval step for all outbound customer communication
Agent updated a record with wrong dataNo validation on write actions before executionAdd a review step that shows the proposed change before it writes
You cannot tell what the agent did yesterdayNo audit log or log is unstructuredAdd structured logging before adding more capabilities
Agent took an action that cannot be undoneNo rollback path defined for that action typeRemove or gate all irreversible actions until rollback is designed
Team is afraid to give the agent more accessCorrect instinct. The scope and controls are not defined well enoughDocument the scope and control layers before expanding permissions
Agent looped or retried an action multiple timesNo idempotency check or retry limit on repeated actionsAdd a max-retry limit and check whether the action already completed before retrying

FAQ

Questions Teams Ask About Agent Control

Does every agent need approval gates?

Level 1 actions (reading, drafting, summarising) do not need gates. Level 2 and 3 actions do. The key question is: if this action runs incorrectly, what is the cost to reverse it?

How do we handle approval without slowing everything down?

Approval gates only slow down high-risk actions, which is the point. Low-risk actions run without gates. Most production agents have 2 to 4 actions that genuinely need approval. The rest can run freely.

What is the right approval channel?

Slack, email, or a simple web interface work fine. The requirement is that a named person sees the proposed action with enough context to approve or reject it in under 30 seconds.

Can we add control layers after the agent is live?

Yes, and it is worth doing even on live agents. Add logging first. It costs least and gives you immediate visibility. Add approval gates next for the highest-risk action types.

How does this relate to autonomous agent architecture?

Control layers are part of the architecture, not an add-on. An agent designed with defined scope, approval gates, audit logs, and rollback paths from the start is far more stable than one that has controls bolted on later.

What about multi-agent systems?

The same rules apply to each agent in the system. In a multi-agent setup, each agent should have its own defined scope. Agents should not inherit permissions from each other by default.

Related Problems

Other Problems Founders Check Around This

Agent Architecture

The scope and the controls should be designed before the agent runs.

If your agent can take actions your team cannot fully explain or audit, the architecture needs review before it gets more capabilities or more access.

Ask About The Architecture