Your agent doesn't need all 58 tools in context. Defer them.
SummaryA thread calling Anthropic's tool-use write-up a new paper is going around X this weekend. It is a year old, and the advice still holds: mark most tools defer_loading and let Claude search for the few it needs.
· MODEL Editors · 2 min read
Some links may be affiliate links. Disclosure

A thread calling Anthropic's tool-use write-up a "paper" is circulating on X this weekend. It is not new. Anthropic published Introducing advanced tool use (opens in a new tab) on its engineering blog on Nov. 24, 2025. The main tip is still worth using if you build agents on the Claude API, and it takes one field per tool.
The problem. Every tool definition you send sits in the model's context before it reads the request. Anthropic's example: five MCP servers (GitHub, Slack, Sentry, Grafana, Splunk) add up to 58 tools and about 55,000 tokens. Anthropic says it has seen tool definitions take 134,000 tokens internally before it optimized them. The docs (opens in a new tab) add that Claude gets worse at picking the right tool once it has more than 30 to 50 to choose from.
The fix. Add a tool search tool to your tools list (tool_search_tool_regex_20251119 or tool_search_tool_bm25_20251119) and set defer_loading: true on the tools Claude doesn't need every time. Deferred tools stay out of context. When Claude needs one, it searches, and the API loads only the matches, five by default. For MCP servers you can defer a whole server and keep one or two of its tools loaded.
The numbers, from Anthropic's internal tests.
- Context used before any work: about 77,000 tokens with every tool loaded, about 8,700 with tool search. Anthropic calls that an 85% reduction.
- MCP evaluation accuracy: Claude Opus 4 went from 49% to 74%, and Opus 4.5 from 79.5% to 88.1%.
- A separate feature in the same post, tool use examples (the
input_examplesfield), raised accuracy on complex parameters from 72% to 90%.
How to set it up. Keep your three to five most-used tools loaded and defer the rest. Name tools by service (github_, slack_) so one search finds the group. Write descriptions with the words people actually use for the task. Add a line to your system prompt listing which kinds of tools exist.
The catch. Tool search adds a search step before each new tool, so it adds latency. Anthropic's docs say to skip it if you have fewer than 10 tools, use every tool on every request, or your definitions are tiny. The accuracy figures are Anthropic's own tests on older models, not independent benchmarks. You still send every tool definition with each request; deferring changes what goes into context, not what you send.
Our take. This is the cheapest agent fix we know of: one field per tool, and no model change. If your agent has a long tool list and keeps picking the wrong one, try this before switching to a bigger model.
Related: Claude Code's Code Mode runs tool calls from JavaScript.
From Anthropic's engineering blog (Nov. 24, 2025) and the Claude Platform docs, read Nov. 22, 2026. MODEL did not run these tests. Not sponsored. -- MODEL Editors
Get three tested tools every Friday
Free, weekly. Unsubscribe anytime.
Sources
- Introducing advanced tool use on the Claude Developer Platform
Anthropic Engineering (Nov. 24, 2025)
- Tool search tool
Claude Platform docs
- Define tools: input_examples
Claude Platform docs
Keep reading

Hot take
AI agents more than tripled commits. Releases rose 30%.
An NBER working paper that tracked more than 500,000 GitHub developers finds AI coding tools add huge amounts of code, but much less of it ships. It is circulating again this weekend. Our take: the bottleneck is now review, not typing.
· MODEL Editors

Tools
Claude Code's Code Mode reaches the stable channel: what it does, and the catch
Code Mode has Claude write JavaScript that runs several tool calls in one step, including background scripts for up to two hours. A week after launch, the stable channel has caught up.
· MODEL Editors

The test
AWS Emulator 1.0.523, tested: a local AWS that refused what it should, and one old-CLI call that went to the real thing
We ran AWS's free local emulator on Linux with no AWS account and no real keys. S3, SQS and DynamoDB worked, it was up in 0.23 seconds, and it enforced the three permission checks we tried. The catch: AWS CLI v1 sent one SQS call to the real AWS endpoint.
· MODEL Editors

AI News
Gemini 4 Neon is out, and in Antigravity it manages Hydrogen
Google's Nov. 18 model sits between Carbon and Hydrogen at $3/$15 per million tokens. In Antigravity, Orchestrator can now pair Neon as planner with Hydrogen as executor. Not tested by MODEL.
· MODEL Editors
FREE · EVERY WEEK
The top AI tools this week, tested.
Every week MODEL's editors pick the AI tools worth your time, try them hands-on, and tell you what's good, what's not, and who it's for. Free, in your inbox.
Free. Unsubscribe anytime. Our picks are editorial and can't be bought.