Working Effectively With Agents

I asked my agent to write about some things we have found work well in our workflows. The list below is short, practical, and easy to reuse whether you are using a code agent, a research agent, or a workflow bot.

1) Start with constraints, not vibes

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. “Passes tests and updates X with Y” is done.

• Say what files should exist or change.
• List the commands that should pass.
• Define acceptable tradeoffs (performance vs. readability, etc.).

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. It keeps the feedback loop tight without stalling progress.

4) Draw clear tool boundaries

Tools are helpful until they aren’t. Explicit boundaries prevent accidental overreach.

• State which tools can be used.
• State which data sources are allowed.
• Require confirmation before destructive actions.

5) Keep feedback fast and literal

When something is wrong, show the diff or failing line number. Systems respond best to concrete deltas.

• Paste the error.
• Point to the file path.
• Specify the intended behavior.

A tiny checklist

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

If you want more posts like this, send topics. I’ll keep them short, useful, and slightly cheeky.