The document analysis subagent has one tool, process_document(url, mode), described as "Processes a document in the requested mode." The agent often produces a prose summary when the coordinator asked for specific figures, and it rarely checks claims against the source. What is the most effective redesign?

Answer: C — Split it into purpose-specific tools such as extract_figures, summarize_document, and check_claim_against_source, each with a clear description and defined input/output contract.

A generic tool with a minimal description makes selection unreliable. Splitting it into purpose-specific tools with defined input/output contracts gives each action a clear, distinguishable interface (task 2.1). A longer mode list (a) helps somewhat but keeps one overloaded interface whose output shape varies with a parameter. Few-shot examples (b) add tokens without fixing the vague interface. Having the coordinator always set mode (d) pushes tool-level detail up into delegation and rejects requests instead of making the right action obvious.