Working Effectively With Agents

Working with agents will become the new normal. The best thing you can do is get started now. This list is by no means exhaustive and you should be iterating based on your own needs and experiences.

1) Start with constraints

Constraints are the rails. They tell the system what to ignore and how to keep you safe from scope creep.

• Specify what must not change.
• Call out any tool restrictions and environment limits.
• If there is a style guide, paste it or link it.

2) Define done in observable terms

“Make it better” is not done.

• Say what files should exist or change.
• List the commands that should pass.
• Set tradeoffs up front.

3) Separate foreground from background work

If you want to keep moving, split tasks into two buckets:

• Foreground: decisions, approvals, and anything with ambiguity.
• Background: chores, sweeping refactors, data cleanup.

If the system supports background runs, use them.

4) Set hard tool boundaries

Tools help until they overreach

• Specify allowed tools.
• Specify allowed data sources.
• Require confirmation for destructive actions.

5) Keep feedback fast and literal

Agents respond to specifics.

• Paste the exact error.
• Point to the file.
• State the exact behaviour.

A tiny checklist

1. Constraints written.
2. Definition of done visible.
3. Tool boundaries explicit.
4. First task scoped to one change.