Tool Search
also called Deferred tool loading, Tool namespacing
Load tool definitions on demand instead of handing over all of them.
A mechanism where the model is given an index of available tools and fetches the full definitions only for the ones it needs, rather than receiving every schema in the request. Keeps large tool surfaces affordable.
The problem it solves is real and gets worse with scale: a platform with two hundred operations cannot put two hundred schemas in every request without spending most of the window on tool descriptions the task will never use.
There is a second, larger benefit beyond the context saving: action selection gets easier. Choosing among two hundred schemas is a genuinely harder problem for the model than choosing among three, so reducing what it sees improves accuracy as well as cost.
The platform that got more reliable as it grew
An internal API surface grows past a hundred operations and tool-use accuracy falls.
Expose a searchable index rather than every schema. The agent loads a handful per task, and accuracy stops degrading with the size of the API — which it otherwise would, since the decision gets harder as the surface grows.
- docsTool search ↗
Deferred tool loading and namespaces: the model searches for tool definitions instead of being handed every schema up front.
- blogCode execution with MCP: building more efficient AI agents ↗
Presenting tools as a code API so the agent loads only what it needs and filters results in the sandbox instead of pushing large intermediate data through the context window.
Tool
A function the model is allowed to call to affect the world.
MCP · Model Context Protocol
An open protocol for plugging tools and data into any agent.
Progressive Disclosure
Reveal detail only when it becomes relevant.
Context Pointer · Reference
A cheap reference to something the agent can fetch if needed.
Context Engineering
Deciding what goes in the window, in what order, at all.