Skip to content

Four role cards: researcher, planner, builder and tester or reviewer. The slide reads: Defined jobs. Allowed access. Visible work.

Agents are teammates with a job, not magic

A useful agent has a clear job, a bounded role and a way to show its work.

Imagine you are joining a project and the team introduces itself. Not with job titles on a wall, but with clear roles and a shared goal. That is a much better picture of an AI agent than a magic button.

Meet the team

  • 01 Researcher. Gathers approved context and surfaces what matters.
  • 02 Planner. Turns a goal into a clear, reviewable set of steps.
  • 03 Builder. Drafts a first version you can react to.
  • 04 Tester or reviewer. Asks what breaks, and for whom. What happens for someone on a slow connection, or with different needs?

Three words that matter

Defined jobs. Allowed access. Visible work.

Defined, because a vague agent produces vague results. Allowed, because good agents have limited responsibilities and limited access. Visible, because you should always be able to see what an agent did and why.

If you can answer the questions what is its job, what can it touch and how will I check its work?, you are working with a teammate. If you cannot, you are working with a mystery, and mysteries are a risky thing to ship.