Design around the work.
Then choose the team.

Agent engineering is the work of defining what an AI agent should do, what it needs to know, how it hands work over and how you decide whether it helped.

1. Choose a repeated task with a visible output

A weekly research brief is easier to evaluate than “be our CEO.” Name the input, the deliverable, the person who reviews it and the condition that means the task is finished.

2. Start with the smallest useful system

One agent may be enough. Add a separate role when it needs different context, tools or evaluation. More agents also introduce more handoffs to supervise. Anthropic’s practical guidance recommends simple, composable patterns before unnecessary complexity. Read the source ↗

3. Write the contract before the personality

Specify purpose, allowed sources, required output, boundaries, uncertainty handling and escalation. Give the agent a concrete first task. A memorable role name can help navigation, but cannot replace an operational contract.

4. Make every handoff explicit

Say who receives the work, what information travels with it and who accepts it. Set a retry or iteration limit. Keep consequential decisions with a named human.

5. Measure a small pilot

SMART means specific, measurable, achievable, relevant and time-bound. If you do not know your baseline, collect it first. “Track time spent creating three briefs this week” is a useful first goal. “Save 80%” without measurements is a guess.

What the planner gives you

Your confirmed business context becomes a recommended architecture, setup prompts, first tasks, metrics and a pilot plan. You review the assumptions and decide what goes live.