Designing tools that make agents useful
The fastest way to make an agent disappointing is to give it a vague mission and a toolbox full of clever actions. The better approach is less magical: define the job, find the decisions inside it, and give the agent a small set of reliable moves.
Start with a job someone already understands
“Manage content” is not a job. “Turn a research packet into a brief with cited claims and three open questions” is. The second version has a clear beginning, a useful artifact at the end, and a person who can tell whether the result is good.
Make every tool boring and explicit
A good tool describes its inputs, validates them, and returns a predictable shape. Prefer search_sources(query, filters) over a generic “browse the web” action. Narrow tools are easier to test, easier to explain, and safer to retry when something fails.
Keep a human in the important loop
Agents should create momentum, not remove judgment. Put review boundaries around claims, external side effects, and anything that changes a customer-facing artifact. The agent can prepare the decision; the person should still own the decision.
A practical starting checklist
- Name the user, trigger, input, output, and review point.
- Give the agent one narrow responsibility before adding autonomy.
- Log tool calls and preserve the intermediate artifacts.