// Before you start
This guide builds on this one:
Work through this before you write code, and again before you ship. Each step is small on purpose.
Define the job#
Write one sentence: "This agent helps who do what." If you need an "and", you probably have two agents.
List the tools, then cut the list#
For each tool, write what it does, what it can change, and what happens if it fails. Remove any tool you cannot describe in one line.
Heads up
Separate tools that read from tools that change things. Reading is safe to retry; changing is not.
Set hard limits#
- Maximum number of steps.
- Maximum cost per run.
- A timeout for every tool call.
Decide what a human sees#
Choose which actions need approval, and make the approval step obvious. It is easier to remove a confirmation later than to add one after an incident.
Write the tests first#
Five tasks with known good answers is enough to start. Run them on every change.