Скачать или смотреть RIP “Dumb” Agents: Why Anthropic’s New Update Changes Everything

  / rip-dumb-agents-why-anthropics-new-update-...  
If you’ve been building AI agents recently, you know the pain. You want an agent that can handle Jira, GitHub, Slack, and your database all at once. But the moment you load those tool definitions, you’ve burned half your context window before the user even says “hello.” Then, when the agent actually runs, it ping-pongs back and forth with the API, eating up tokens and time while “thinking” about every single step.

The era of that “dumb” agent architecture is ending.

Become a member
Anthropic just announced three massive features for the Claude Developer Platform that shift the paradigm from simple function calling to intelligent orchestration. Here is why your next agent will be faster, cheaper, and infinitely smarter.

1. The “Tool Search” Tool: Stop Loading the Whole Factory
Until now, if you wanted an agent to have access to 50 tools, you had to shove all 50 definitions into the system prompt.

The Problem: Internal data shows that just defining tools for a 5-server setup (GitHub, Slack, Sentry, etc.) can burn 55,000 to 134,000 tokens upfront. That’s expensive and leaves no room for actual work.
The Fix: The Tool Search Tool. Instead of loading everything, you mark your heavy tools as defer_loading: true. Claude now acts like a smart developer: it sees a "search" tool, looks for what it needs (e.g., "github tools"), and only then loads the relevant definitions.
The Result: In testing, this reduced context consumption from ~77k tokens down to ~8.7k tokens — preserving 95% of the context window.
2. Programmatic Tool Calling: The “Sandbox” Revolution
This is the game-changer. Traditional agents rely on “natural language tool calling.” They fetch data, read it, decide the next step, and fetch more data.

The Problem: It’s called Context Pollution. If your agent needs to find two employees over budget, it might fetch 2,000 expense lines, read them all (filling your context window), and manually calculate the sum. It’s slow, prone to math errors, and insanely token-heavy.
The Fix: Programmatic Tool Calling. Claude can now write a Python script to run in a secure sandbox. It loops, filters, and calculates outside the context window.
The Result: Using the budget example, Claude writes a script to process the data in the background and returns only the final answer (e.g., “Alice and Bob are over budget”).
Data processed: 200KB of raw data becomes just 1KB of result in context.
Efficiency: Token usage drops by 37% on complex tasks.
3. Tool Use Examples: Show, Don’t Just Tell
JSON schemas are great for structure, but terrible for nuance. They tell the model what a field is, but not how to use it.

The Problem: Ambiguity. If a field is due_date, should the model send "2024-11-06" or "Nov 6"? If the model guesses wrong, the API call fails.
The Fix: Tool Use Examples. You can now embed 3–5 concrete examples of correct usage directly into the tool definition.
The Result: This “few-shot” prompting within the tool definition cleared up confusion on complex parameter handling, boosting accuracy from 72% to 90% in internal tests.
The Bottom Line
We are moving away from agents that blindly flail at APIs. With dynamic discovery (Tool Search), efficient execution (Programmatic Calling), and reliable invocation (Examples), developers can finally build agents that scale without breaking the bank.

These features are available in beta right now. It’s time to upgrade your agents.

Source Reference: Introducing advanced tool use on the Claude Developer Platform. Anthropic. Available at: https://www.anthropic.com.