CCA-F start here · Trap gym · Too many tools

Too many tools

A multi-agent system has 18 tools, and every agent was given all of them. Picks get worse, and agents reach for tools that belong to someone else's job. The tempting answers polish the tool list: better descriptions, a prompt that says “don't use these”, a router. The exam's answer is to hand out fewer tools.

D2 Tool Design & MCPExam guide 2.3About 15 minutes

At a glance: give each agent the 4 or 5 tools its job needs, not all 18, with a small shared tool only for a need that comes up often.

Start here: the chef's counter

Picture two kitchens.

A hundred gadgets on every counter. Each chef spends time choosing and sometimes grabs the wrong one. And because the pastry chef's counter has a meat cleaver too, one day it gets used on a cake.
Five knives per station. Each chef has exactly what their dishes need, works fast and picks right. If the pastry chef often needs a pinch of salt, a small salt pot sits on that station. Anything bigger goes through the head chef.

A multi-agent Claude system works the same way. A coordinator (the head chef) hands work to subagents (the stations). Each subagent's tool list is set in its AgentDefinition, the record that gives a subagent its description, its system prompt and the tools it may use. Claude picks a tool by reading the descriptions of every tool it holds, so every extra tool is one more wrong answer it could choose. A tool from another role's job is worse: the agent may use it to do that job, badly and unchecked.

Coordinator splits the work, routes hard questions Search finds papers, trials 4 tools Analysis reads, extracts 5 tools Synthesis writes the draft 4 tools Reporting formats, exports 5 tools

The diagram follows the simulator: the counts change with how you hand out the tools.

The one question to ask: does this agent really need every tool it has?

Words in the question that tell you which side you're on

Scope the tool set

  • “18 tools”, “every agent gets every tool”
  • “built from one template with the full tool set”
  • “uses a tool outside its role”, “searches in the middle of a draft”
  • “selection got worse after adding tools”

Add one limited shared tool

  • “often needs to look up”, “high-frequency”
  • “most checks are one-field lookups”
  • “keeps handing back to the coordinator for one lookup”
  • “every round trip adds a long wait”

Try it: hand out the tools

A research system has 18 tools across four roles. Pick a request, then pick how the tools are handed out. The grid shows what each agent holds; the replay shows which tool each agent calls at each step.

1. The request
2. How the 18 tools are handed out

What each agent holds

search toolsanalysissynthesisreportingverify_factanother role's tool

One replay of the request

    Selection reliability: the chance an agent picks the right tool at any one step, given how many it holds (illustrative):

    The reliability figures and minutes are illustrative, chosen to show the shape. They are not measurements. The exam guide's point is that 18 tools instead of 4 or 5 degrade selection, that agents misuse tools outside their specialisation, and that a limited cross-role tool suits a high-frequency need.

    Things to try: full review + all 18 to every agent, then + scoped tools to see the misuse disappear. Then 12 checks + scoped tools to watch the round trips pile up, and 12 checks + scoped plus verify_fact. Last, one agent with all 18 on either request.

    The four traps

    TrapWhat the wrong option looks likeWhy it loses
    Only usually works“Rewrite all 18 tool descriptions”; “list in each agent's prompt which tools it must not use”Helps a little. Good labels don't fix the fact that 18 choices degrade selection, and a tool the agent holds is a tool it may still misuse. Remove it instead.
    OverbuiltA router agent that picks tools for the others; a hook that blocks off-role callsWorks, but adds a model call or code to maintain, plus a new failure point, when scoping the list removes the cause.
    Too strictNo cross-role tools at all, so the agent asks the coordinator for every lookupCorrect but slow. For a frequent need the round trips add delay; a limited cross-role tool is the guide's answer.
    Over-provisioned“Give the synthesis agent the search agent's full tool set” to cut round tripsSolves the delay by bringing back the cause: off-role tools it can misuse. Give one narrow tool, not the whole set.

    Practice: three exam-style questions

    Answer first, then open “Think it through”.

    Rule to remember

    Fewer, role-fitted tools beat more tools with better labels. Scope each agent to its job; add one narrow shared tool only for a need that comes up often; send hard cross-role questions through the coordinator.

    Exam guide 2.3: “Too many tools (18 vs 4-5) degrades selection reliability.” “Agents with off-specialization tools misuse them.” “Scoped access; limited cross-role tools for high-frequency needs.”

    ← All traps