Definition
Deferred tool loading delays adding a tool's full definition to model context until the tool is discovered or needed. The initial context can provide a compact description of the available capabilities, with detailed schemas loaded later.
This reduces the up-front context devoted to a large tool catalog. It also introduces a discovery step: a tool cannot help if the system fails to expose or find it at the moment it is needed.
Origin and current usage
Anthropic's November 2025 Tool Search Tool announcement described marking tools for deferred loading and expanding matching references into full definitions after discovery. It is an earlier documented implementation of the mechanism. The announcement does not establish who first coined deferred tool loading.
OpenAI's tool-search documentation describes searching for deferred tools and loading their definitions at runtime. Namespaces and MCP servers can initially expose broader descriptions.
Earendil listed deferred tool loading among Pi 1.0's October 1, 2026 additions.
Scope and implementation differences
Providers defer different parts of a catalog and expose different search interfaces. OpenAI can retain a function's name and description while deferring its parameter schema; Anthropic's announcement describes loading matching tools after search. Tool search is one discovery mechanism, while deferred loading describes when the definitions enter context.
Operational significance
Measure discovery success as well as context savings. A concise catalog can reduce input size yet hide a rare capability, confuse similar tools, or require extra searches. Evaluate realistic requests that need infrequently used tools and inspect whether the correct definition is loaded.
Loading a definition does not grant authorization. The application must still enforce tool access, validate arguments, and check the permitted effects when execution occurs. Treat loaded schemas and descriptions according to their source and trust level.
Cache behavior also varies by implementation. Adding definitions at a compatible point can preserve a reusable prefix; that does not guarantee a cache hit or a cheaper run for every request.
Distinguish it from nearby terms
- Tool search locates relevant capabilities. Deferred loading governs when their full definitions enter context; another discovery mechanism could serve that role.
- Tool routing chooses which tool to use. Loading makes a tool describable or callable but does not make that selection correct.
- Prompt compression reduces an existing input. Deferral avoids adding some input until it is needed.
- Programmatic tool calling coordinates invocations using code. Tools may be loaded eagerly or deferred in either calling style.
Check your understanding
A deferred catalog saves thousands of input tokens but causes the agent to miss the rollback tool. Which discovery tests should fail, and why does the smaller prompt alone not establish an improvement?