# How do I make my AI agent ask before it makes a decision?

_Make agents pause for human decisions by defining approval boundaries, structured options and a durable inbox outside the chat session._

By Saad Ahmed, 2026-10-20

To make an AI agent ask before deciding, define which choices require human approval and give the agent a durable place to pause and present options. The agent should record the question, recommendation, consequences and blocked work, then wait for an authenticated answer.

“Ask when unsure” is too vague. Models can be confident when a decision is subjective, expensive or difficult to reverse.

## Define approval boundaries in advance

Require approval for decisions such as:

- spending money or changing a budget;
- deleting or irreversibly overwriting data;
- merging, deploying or publishing when not already authorized;
- expanding credential or permission scope;
- choosing product policy, legal language or brand claims;
- changing the accepted outcome or scope of a task.

Let the agent handle reversible implementation details inside an approved brief. Too many approval gates turn the person into a command router; too few let the agent make business decisions it does not own.

## Ask a question a person can answer

A useful decision request contains one clear question, two or three concrete options, and a recommendation with a short reason. It also says what work is blocked and what can continue safely.

Do not bury the decision beneath logs. Keep test output and implementation detail in linked evidence so the approval card remains understandable.

## A file or issue-tracker method

Add an `Approval required` section to the task. When the boundary is reached, the agent stops the affected action, writes the options and changes the status to waiting for human. The person records the choice in the same task before work resumes.

This survives a chat session, but notifications and answer identity depend on the tools you already use.

## Where ecco fits

ecco separates agent progress from decisions that need a person. An MCP-connected agent can raise a structured decision on the task; the person sees it in the inbox, chooses an option, and leaves a durable answer for the current or next agent session.

The board records the choice, but it does not widen the agent's permissions. Approval remains specific to the stated decision and task.

[Connect an agent to ecco](/blog/connect-an-agent), or compare this shared workflow with [Claude and ChatGPT used on their own](/blog/ecco-vs-claude-and-chatgpt).

## Frequently asked questions

### What decisions should an AI agent always ask about?

Require a person for irreversible actions, money, credentials, external publication, legal or brand commitments and changes to the agreed outcome. Adjust the list to your risk level.

### Should an agent stop all work while waiting?

Only the affected action must stop. The agent may continue clearly independent, reversible work if the task rules allow it and doing so will not prejudge the decision.

### How should the agent present options?

Put the question first, keep the options mutually exclusive, state the main consequence of each and identify the recommended option. A person should be able to answer without reading a full transcript.

### Can a previous approval apply to a new task?

Do not assume it does. Record approvals with their task, scope and context; ask again when the material facts or requested permission change.