Recommended Free Tools
If your agent sees every tool definition before it knows what a request needs, a large catalog can consume context on tools it never uses. Keep frequently used tools eager; defer less common ones behind a discovery step when the catalog’s context cost justifies the added retrieval and latency. There is no universal tool-count cutoff: choose by testing task success, tool selection, input-token use, and end-to-end time on representative requests.
What eager and deferred loading mean
With eager loading, the agent receives the full definitions for the configured tools up front. This is simple and can work well when the catalog is small or most tasks use most of its tools. The cost is that definitions the agent does not need still occupy initial context.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
With deferred loading, only selected tools are initially exposed. The agent can search for other tools when a request calls for them; matching tools are then made available with their full definitions. This reduces initial definition context but adds a discovery step, and the result depends on the agent being able to find the right tools from their names and descriptions.
When to use each approach
| Consideration | Eager loading | Deferred loading |
|---|---|---|
| Initial context | Includes all configured tool definitions, including those unused for a particular request. | Includes the search mechanism and initially exposed definitions; matching deferred definitions are loaded when needed. |
| Tool selection | The agent can see all definitions at once. | Selection depends on discovery and the clarity of tool descriptions. |
| Latency | Avoids a tool-discovery step. | Discovery can add a step; measure its effect in the full request, not just the search call. |
| Implementation | Straightforward for a small, stable toolset. | Requires configuring deferred definitions and a supported search flow. |
| Best fit | Small catalogs, or tools used in most requests. | Large catalogs with distinct tool groups and meaningful unused-definition overhead. |
Anthropic suggests considering tool search when a toolset grows past roughly 20 tools or baseline context use becomes noticeable. That is provider guidance, not a universal threshold. Anthropic also characterizes a typical multiserver setup as having about 55,000 tokens of definition overhead, and says its described tool-search setup typically reduces tool-definition context by over 85% by loading the 3–5 tools needed for a request. These are Anthropic’s figures, not independent measurements or guaranteed savings; actual results depend on the catalog and workload. See its tool-context guidance.
#1 Best Overall
How to configure deferred tools
OpenAI Responses API
OpenAI’s Responses API supports tool search for deferred function and MCP-server tool definitions. Enable the tool-search flow and mark the definitions intended for deferred loading accordingly. Consult the current OpenAI tool-search guide for the supported configuration and request format.
OpenAI Agents API
Function definitions load eagerly by default in the Agents API. To defer selected functions, configure tool search and set defer_loading: true on those definitions. OpenAI recommends clear names and descriptions; grouping deferred functions into namespaces or MCP servers can also help organize discovery. The exact supported setup is described in the OpenAI tool-search guide.
OpenAI MCP tools
OpenAI documents tool search for functions exposed by an MCP server and says it is most useful when a server exposes many functions. Use the documented tool-search flow for the MCP tools you want deferred; do not assume that connecting a server alone means every provider will discover its tools in the same way. See OpenAI’s MCP server documentation.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Anthropic
Anthropic’s tool-search flow lets you provide tool definitions and mark tools for deferral with defer_loading: true, while keeping a search tool available. Search results expand matching tools into full definitions. Follow the current Anthropic tool-search documentation for the precise format and supported behavior.
Design the catalog so discovery works
Deferral shifts some work from reading every definition to finding the right definition on demand. Make each tool easy to distinguish from its neighbors.
- Use names that describe the action or resource, rather than internal abbreviations.
- Write descriptions that say what the tool does, when to use it, and any important scope or limitations.
- Group related deferred tools into coherent namespaces or MCP servers where the platform supports it.
- Keep tools eager when they are needed for most requests or are essential to a reliable first step.
Test the choice on real workloads
Compare eager and deferred configurations using the same representative requests and measure the whole task. A lower initial token count is useful only if the agent still finds the right tools and completes work reliably.
- Choose representative tasks. Include common requests, requests that need several tools, and cases where similar tools could be confused.
- Run both configurations. Keep prompts, tool behavior, and evaluation criteria consistent so the loading strategy is the meaningful difference.
- Record the outcomes. Track task completion, whether the agent selected the appropriate tools, input-token use, and end-to-end latency including discovery.
- Inspect failures. If deferred tools are missed or confused, improve descriptions and grouping, or keep those tools eager. If savings are negligible or the added step slows common tasks, simplify the catalog or retain eager loading.
- Recheck after changes. Repeat the comparison when the catalog, task mix, or provider configuration changes.
OpenAI likewise recommends evaluating representative requests by task completion, input-token usage, and latency in its tool-search guidance. The official documentation describes implementation and provider-reported results; it does not establish a controlled, independent head-to-head winner across providers.
Account for other context-management options
Tool search addresses the context taken by tool definitions. It is not the only way to manage context: Anthropic discusses prompt caching and context editing as complementary approaches in its context-management guide. Which combination helps depends on the runtime and workload, so evaluate the complete request path rather than assuming that deferral alone solves context pressure.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
