Measure whether the bounded research agent works

Translate the claim into assertions or a tiny eval harness. The metric to protect is: step count ≤ max; citation coverage; abstention on empty search. Always record the denominator (how many cases) beside any rate. A percentage without a denominator is marketing, not measurement.

const parsed = OutputSchema.safeParse(raw);
if (!parsed.success) {
  throw new ContractError("invalid_output", parsed.error.issues.map(i => ({
    path: i.path.join("."), code: i.code
  })));
}
return parsed.data;

Expected evidence: The agent searches once, cites kb://returns, and stops before MAX_AGENT_STEPS.. A passing assertion proves only the behavior it names; broader usefulness still needs the chapter’s full limits.

Be explicit: beating the baseline (single-shot LLM answer without search) on this fixture does not prove behavior under infinite tool loop, or final answer with no citations after search. Label observations separately from conclusions so the next page inherits honest evidence about the bounded research agent.

Compute step count ≤ max; citation coverage; abstention on empty search with the denominator written beside the rate every time. For this chapter, the evaluation set is intentionally tiny; that is allowed only if you say so in the evidence. Compare against single-shot LLM answer without search before celebrating.

Add one negative case aimed at infinite tool loop, or final answer with no citations after search. A suite with only happy cases cannot protect the bounded research agent when the characteristic failure appears in review.

If a check is expensive or flaky, shrink it until it is deterministic on fixture KB with 4 notes + max_steps=3. Flaky green builds teach the team to ignore gates. Record what this page does not prove so security-ops and mastery-ship inherit honest limits.

At the evaluation stage for agents-in-code, the job is narrower than finishing a product demo. You are creating one progressive evidence piece about fixture KB with 4 notes + max_steps=3 that later pages inherit without redefining success. Keep that fixture small enough to inspect by hand, keep outputs copy-pasteable as text, and refuse to narrate this baseline as if it were a production SLA: single-shot LLM answer without search.

For this page specifically, success looks like metrics with explicit denominators and a negative case while still centering the user decision to search a local knowledge base then produce a cited answer within a step budget. If you cannot point to a file, command, or assertion that proves that for the bounded research agent, stay on this page instead of advancing.

Glossary: tool · Glossary: structured output · Cheatsheet: production ops signals

Previous · Next