passdrill
AI & LLM Engineering · AI Agents & Tool Use · Card 007/012 medium

A long-horizon research agent needs to answer a multi-part question that requires several separate lookups. One design has the model produce an upfront plan listing each lookup it will need and what each one depends on, hand that whole plan to a separate component that executes each lookup, and only then pass the collected results to a final step that composes the answer -- rather than re-invoking the full reasoning model after every single lookup to decide the next step. What is the main advantage of this plan-first, decoupled design over re-reasoning after every individual lookup?

  1. It is the only design capable of using more than one external tool within a single task
  2. It guarantees a higher final accuracy than any design that reasons between each lookup, on every possible task
  3. It eliminates the possibility of any lookup ever failing or returning an unusable result
  4. It avoids repeatedly re-feeding the full growing transcript of prior reasoning and results back into the expensive reasoning model at every single step, cutting redundant token consumption while remaining robust even if an individual lookup tool fails
Next card → Shuffle