Michele Meloni
Contact

Write to me directly.

For a collaboration, a question about my work or an initial conversation, use this form. For a detailed estimate, use the dedicated flow on the Services page.

Your message is sent directly to info@michelemeloni.com.

Back to articles
Article · AI & Product DesignMichele Meloni8 min read

Designing experiences with AI agents

How to give a system autonomy without taking away people’s understanding, control or ability to stop and correct its actions.

Published · 4 September 2026Updated · 4 September 2026

00 · Premise

The more a system can act, the clearer its boundaries and consequences must be.

An agent does more than respond: it interprets a goal, selects tools and executes steps. The experience must make clear what it will do, which data it will use, when it will ask for confirmation and how control can return to the user.

Governable autonomy

Goal, proposal, confirmation and execution.

A conversation can remain simple even when the system coordinates many steps. Operational complexity should become visible only when it helps someone understand a decision, authorise an action or verify the outcome.

  1. 01Goal

    The system confirms what it understood.

  2. 02Proposal

    It shows the plan, tools and data involved.

  3. 03Confirmation

    It asks before sensitive actions.

  4. 04Execution

    It exposes progress, outcome and recovery.

Promptario interface with resources for artificial-intelligence workflows
Promptario · organise tools and methods before entrusting them with a task.
01

Choose work that genuinely needs an agent

Autonomy is useful in journeys with multiple steps, exceptions or unstructured information. For simple deterministic actions, a conventional interface may be clearer and more dependable.

Practical actions

  • Describe the outcome the agent must produce.
  • Check whether fixed rules would solve the problem better.
  • Define which decisions must never be delegated.
02

Make capabilities and permissions explicit

People need to know what the agent can read, change or send. Permissions should match the task and be requested when they become necessary.

Practical actions

  • Describe tools and data in plain language.
  • Separate reading, proposing and executing.
  • Avoid broad consent for undefined future actions.
03

Match confirmation to risk

Confirming every step creates friction; never confirming makes the system opaque. Control should increase with cost, irreversibility and external impact.

Practical actions

  • Automate only low-risk, recoverable actions by default.
  • Show a summary before sending, purchasing or publishing.
  • Allow parameters to be corrected without restarting.
04

Keep status and stopping controls visible

A long task should communicate what is happening, which steps are complete and when attention is needed. Stopping and resuming belong in the journey.

Practical actions

  • Show progress and the last verified result.
  • Provide stop, undo and resume controls.
  • Return control to a person after repeated failures or before high-risk actions.

Ready to use

Before enabling an agent

  • The use case genuinely benefits from autonomy.
  • Capabilities, data and limits are understandable.
  • Permissions are minimal and contextual.
  • Sensitive actions require confirmation.
  • Status, stop, undo and recovery are designed.
  • A human handoff exists when needed.

Key principle

Autonomy should increase effectiveness, not reduce control.

A credible agent makes its work understandable and gives people a real ability to authorise, correct or stop it.