AI agents shouldn't wait for instructions

Most systems people call "agents" are still chatbots. They sit in a text box and wait. A human types, the model answers, maybe it calls a tool, and then it waits again.

That is assistance. It is useful, but it does not remove work. Every step still needs a person to decide the next step.

An agent that is worth the name has three things: a goal, a loop and permission to act.

A goal, not a prompt

A prompt describes an answer. A goal describes an outcome.

"Summarize this ticket" is a prompt. "Close this ticket" is a goal. The second one implies research, a decision, an action and a check. It also implies the agent might come back and say the ticket cannot be closed, and why.

When I design an agent, the first artifact is not the system prompt. It is a one-line definition of done that a machine can verify.

A loop, not a turn

The chat pattern is one turn: input, output. Agents run in a loop: observe, decide, act, check, repeat until done or blocked.

while not done:
    state   = observe()
    action  = decide(state, goal)
    result  = act(action)
    done    = verify(result, goal)

The loop is where autonomy lives. It is also where things go wrong, so it needs limits: a maximum number of steps, a budget, and a list of actions that require a human.

Permission to act

This is the part teams avoid. An agent that can only read is safe and useless. An agent that can write, deploy, send and publish is useful and scary.

The answer is not to remove permissions. It is to make them explicit and tiered:

• Free actions. Read, search, draft, run tests. No approval.
• Reviewed actions. Open a pull request, propose a reply. A human sees the result before it ships.
• Gated actions. Deploy to production, send money, delete data. A human approves each one.

Most of the value sits in the first two tiers. Once they run without supervision, the third tier gets smaller than people expect.

AI should reduce the number of decisions humans need to make.

What changes

When agents stop waiting, the shape of the work changes. People stop supervising steps and start reviewing outcomes. Backlogs get processed overnight. The job becomes writing better goals and better checks.

That is the system I want to build for every client: not a smarter assistant, but fewer decisions on a human's desk.