Agent Routing

How models should choose Fable tools and avoid inventing capabilities.

Agent routing is the process of deciding whether a prompt should be answered in text or with a Fable UI tool call.

Today

Today the model chooses from tools registered in toolRegistry. The current registry exposes only show_metric.

The AI SDK route passes those tools to streamText with toolChoice: "auto". The model decides whether to call a tool based on the system prompt, the tool description, and the conversation.

Tool descriptions must be short and constrained. show_metric says to use it for a single total, count, amount, percentage, status number, or KPI. It also says not to use it for lists, tables, browsing, forms, suggestions, or actions.

That boundary is important. If a prompt asks for "the last 10 orders", the model should not squeeze the answer into a metric card.

Manifests

Manifests give routing more structure than a short tool description can. A manifest should include:

Use-when rules.
Do-not-use-when rules.
Neighboring tools.
Positive trigger prompts.
Anti-trigger prompts.
Conflict prompts.
Safety notes.

The MetricCard manifest already describes the boundary around show_metric. Automated manifest eval execution is planned, not shipped.

Constraints

The model must not invent:

Resource IDs.
Component names.
SQL.
Firestore queries.
Backend operations.
Raw endpoints.
Authorization decisions.

The model selects from allowlisted capabilities. The host app validates everything that matters.

Future resources

Future data-backed components should expose resources through a safe agent resource catalog. The catalog can include labels, descriptions, aliases, allowed filters, allowed actions, and routing hints.

It must not expose tokens, headers, raw endpoints, collection paths, handler functions, or private auth logic.

Testing routing

Routing should be tested with:

Positive prompts that should select a tool.
Anti-triggers that should not select that tool.
Conflict prompts where a neighboring tool is better.

For show_metric, "What is today's revenue?" is positive. "Show the last 10 orders" is an anti-trigger. "Show top selling items today" is a conflict prompt that should prefer a future table or browser surface.