Tips for Using Agents

Agents work best when the task is concrete. Ask for a specific deliverable, define the scope, and make the stopping condition explicit. "Compare these three papers and extract conflicting findings" is much stronger than "look into this topic."

Give the right context

Provide context in the order a human would need it:
  1. Objective
  2. Constraints
  3. Inputs
  4. Output format
If the task is high stakes, say so. If a source is authoritative, point to it. If you want a short answer, citations, or a table, specify that at the start.

Ask for verification

Ask the agent to show assumptions, uncertainties, and how it checked the answer. This usually produces more reliable results than asking whether it is confident.
Good pattern:
  • "Do the task, then list assumptions, open questions, and how you verified the answer."

Use iteration

Treat the first answer as a working draft. Agents are usually strongest in short loops:
  1. Draft
  2. Critique
  3. Refine
  4. Verify
This is especially useful for research, analysis, and coding tasks.

Set boundaries

If the agent can browse, run code, or edit files, define the allowed surface area. Clear permissions and constraints reduce irrelevant work and risky actions.
For longer tasks, add checkpoints. For example, ask the agent to pause after planning, after evidence gathering, or before destructive changes.

Make outputs inspectable

Ask for outputs you can audit quickly. Tables, direct quotes, file paths, citations, and explicit calculations are easier to verify than polished summaries.
If you cannot inspect the result quickly, the workflow is probably too opaque.

Use agents where they help most

Agents are most useful when work involves multiple steps, many sources, or repetitive reasoning. For simple facts or trivial edits, the overhead can outweigh the benefit.