Tool distribution and tool choice

Tool count is a reliability factor

Give an agent eighteen tools where five would do and selection reliability degrades. Every additional tool makes the selection decision more complex.

Each additional tool gives the model another possible choice. When several tools have similar purposes, adding more can increase selection errors.

Agents misuse tools outside their specialization

An agent may use any tool it receives, even when the tool does not suit its role. For example, a synthesis agent with a web-search tool may start new searches instead of combining the research it was given.

Scope each agent's tool set to its role. The restriction is not bureaucratic; it prevents a whole class of misuse.

Remove tools a role should not use, but keep the tools it needs to finish its work. For example, a triage agent may need both a classification tool and a way to route a contact it cannot classify. The goal is the right set of tools, not the smallest possible set.

Narrow a tool when the role still needs it

Sometimes a role genuinely needs a capability but not in its general form. Replace a generic fetch_url with load_document, which validates that the URL is a document. Same need served, far less room for misuse.

For a common need shared across roles, provide one narrowly scoped tool. For example, a synthesis agent might receive verify_fact. Send more complex cases to the coordinator instead of giving every agent more tools.

tool_choice

• auto — may call a tool or answer in text.
• any — must call a tool, chooses which. Use it to guarantee the model acts rather than returning conversational text.
• Forced, {"type": "tool", "name": "..."} — a specific tool must be called. Use it to guarantee ordering, such as running extract_metadata before any enrichment tool, then continuing in follow-up turns.

The distinction to hold: any guarantees that a tool is used. Forcing guarantees which.

In some card games you have to play a card every turn, even when passing would serve you better, so a player with nothing useful still puts something down. Requiring a move is not the same as requiring a particular one.

The three values as they are actually written:

{"type": "auto"}
{"type": "any"}
{"type": "tool", "name": "extract_metadata"}

The third differs from the second by one field, and that one field is the entire difference between "act" and "run this".

There's a cost to any worth knowing. It forbids a text-only reply, so an agent with the work finished has no way to say so and no way to ask you a clarifying question; it calls something instead. Set it on the turn that needs action, then hand the turn back.