Getting started

Five tips for your first agent project.

Five things that separate agent projects that ship from ones that do not

These are, in effect, the lessons of every early agent deployment. None of them are novel — all of them are repeatedly violated.

1. Start narrow, not broad

Pick one high-friction task — summarising meetings, triaging customer emails, researching competitors — and build an agent that does only that. A Task Agent that does one thing well beats a Strategic Agent that does ten things badly.

Narrow scope compresses the evaluation problem, the governance problem, and the stakeholder conversation. It also means you can ship.

2. Define clear boundaries

Write the NOT for list before you write anything else. Specify what the agent can't access, when it must escalate, and who approves which actions. Boundaries go into the system prompt, into tool wiring, and into REMIT — all three.

3. Use real-world workflows, not demos

Map the agent into a workflow that already exists. If you're replacing a manual step, replace exactly that step. Do not imagine a greenfield workflow where the agent is central — that workflow will never exist.

4. Focus on feedback loops early

From day one: capture user feedback, error reports, and learning signals. This becomes your silver dataset, which you promote to gold via human review, which becomes your eval set. No feedback loop means no learning, means no trust, means no graduation up the authority levels.

5. Choose tools that let you iterate fast

Use platforms that support modular iteration and debugging: Claude's Projects and Skills, the AI Gateway for model switching, an eval platform from day one. The cheapest evaluation is the one you can run in 30 seconds.