> ## Documentation Index
> Fetch the complete documentation index at: https://hastekit.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Deferred Tools

> Load tool definitions on demand so big tool catalogs don't bloat the context window

Modern agents often have access to dozens or hundreds of tools — between MCP servers, OpenAPI specs, and connectors, the catalog grows fast. But every tool the agent *might* call eats prompt tokens, every turn, whether the agent uses it or not. **Deferred Tools** fix that: tool definitions stay out of the system prompt until the agent explicitly asks for them.

<Frame>
  <img src="https://mintcdn.com/hastekit/890s5wy2NHvxQLSN/images/agent-builder/deferred-tools.png?fit=max&auto=format&n=890s5wy2NHvxQLSN&q=85&s=030a71975ef5064abd3685be4487aebd" alt="Deferred tool toggle" width="3886" height="1634" data-path="images/agent-builder/deferred-tools.png" />
</Frame>

## How it works

A deferred tool is registered with HasteKit but not exposed in the system prompt by default. Instead, the agent sees a single meta-tool — `tool_search` — that lets it discover and load tools on demand:

1. The agent calls `tool_search("calendar")`.
2. HasteKit returns a short list of matching tool names and one-line summaries.
3. The agent calls `tool_load(["GoogleCalendar.CreateEvent"])`.
4. The full schema of that tool is now in scope and the agent can call it.

The discovery + load step costs one or two extra LLM turns up front, but means the agent isn't paying for tool definitions it never uses — and you can attach huge tool catalogs without context bloat.

## Marking tools as deferred

Available on every tool type that supports filtering:

* [MCP Servers](/docs/gateway/agent-builder/mcp-server) — **Deferred Tools** field (comma-separated tool names).
* [API Servers](/docs/gateway/agent-builder/api-servers) — flag specific operations.
* [Connectors](/docs/gateway/agent-builder/connectors) — flag specific actions.
* [Provider Tools](/docs/gateway/agent-builder/provider-tools) — toggle on a per-tool basis.

## When to defer a tool

* **Big catalogs** — anything past \~20 tools starts to crowd the prompt and degrade tool-call accuracy.
* **Specialty tools** — tools the agent only needs in narrow cases (e.g. `Jira.TransitionIssue` is only relevant once the user mentions tickets).
* **Heavy schemas** — tools whose input schemas are huge (deeply nested objects, long enums) cost more tokens to keep loaded.

## When NOT to defer

* **Always-needed tools** — if the agent is going to use the tool on most runs, deferring just costs an extra round-trip every time.
* **Tools the agent should consider proactively** — discovery only happens when the agent thinks to search; if the right answer is "use tool X" and the agent doesn't know X exists, it won't find it.

## Pair with approval

Deferred and **Requires Approval** are independent. A tool can be deferred (not in the prompt), approval-gated (held for human review when called), both, or neither.