Distribute tools appropriately across agents and configure tool choice

Give every agent the full toolset so none of them is ever blocked.

Each additional tool widens the decision, so selection accuracy falls — and agents reach for tools outside their specialisation, like a synthesis agent running its own searches.

Scope tools to the role. A blocked agent is a routing question for the coordinator, not an argument for a bigger toolbox.

Set tool_choice to "any" so the model is guaranteed to pick the right tool.

On configurations that support forcing, "any" requires some tool request, not a specific one. Models that reject forcing require a different integration.

Check model compatibility: "any" requires some request and named selection constrains its identity. The application owns authorization and execution order.

Force the required first tool and let the model handle the rest in the same turn.

Forcing applies to the request you send it on, so the ordering guarantee does not extend across the steps that follow.

On a supported model, require the first tool request, validate and execute it, then expose the next step. With auto, inspect the response before advancing.

Route every cross-role need back through the coordinator, for cleanliness.

A high-frequency check becomes a round trip through the hub every time, which dominates latency for something narrow and well-defined.

Give a narrow scoped tool for the frequent case; route the complex ones through the coordinator.

Keep a broad utility tool available because it covers many situations.

A permissive tool accepts inputs it should refuse, so misuse shows up as plausible-looking output rather than an error.

Prefer a constrained tool that validates what it is handed over a general one that will accept anything.

model-specific tool selection support (checked 2026-09-21)

Answer this: Task 2.3 covers auto, any and forced tool selection.

The docs now say: Claude Fable 5.1 rejects any and named selection. Use auto with strict tools and response checks, or structured outputs; selection, schema conformance and execution are separate controls. Read it