Most agent tutorials start general and keep adding tools. For a job you run every week, I've found it works better the other way around: write the one job on a card first, then let the card design the harness.
Six lines: input, steps, output, done when, risks, never. Each one turns into code:
- Steps become tools, one per verb.
- Done when becomes the eval suite.
- Risks become hard caps: searches per turn, max steps, a token budget.
- Never becomes the tools you leave out. No shell, no writes outside one folder. It's the cheapest safety feature there is.
The detail that mattered most for a research agent: models can produce a plausible URL that was never retrieved. So the harness keeps its own ledger of pages actually fetched (search snippets don't count) and refuses to save a brief citing anything else.
In my evals Haiku 5.5 passed 12 of 12 runs at about $0.43 per suite, Sonnet 5.5 passed 6 of 6 at about $2.54. Small sample, but enough to make Haiku my default for routine research.
If you'll only run a workflow twice, a general agent is probably the better use of your afternoon. If it repeats, the card gives you a cost ceiling and real tests.
The full step-by-step build (about 250 lines of Python with the Claude API, no framework) is here: Pick One Workflow, Then Build the Harness Around It
Top comments (0)