Agentic CustomizationThe built-in tool catalogue, and what it costs
No narration yet
Module 3, Lesson 125 min

The built-in tool catalogue, and what it costs

You do not have to build everything. Anthropic ships a catalogue of tools that run without you implementing anything, and Agentic Customization is 4.1% of the paper, which is mostly this list plus knowing what it costs.

The GA list

Code execution. Runs code in a sandbox. GA, despite being the one that sounds like it should still be experimental.

Bash. Shell commands.

Text editor. Structured file viewing and editing, the same shape of capability Claude Code uses.

Memory. Persistence across turns without you hand-rolling a store.

Tool search. For when you have too many tools to put all their definitions in every request, which is a real problem the moment you connect a couple of MCP servers.

Web search. $10 per 1,000 searches, plus the tokens the results burn.

Web fetch. Free, plus tokens.

Computer use is the beta

Computer use is still beta. Beta header computer-use-2025-11-24, and computer-use-2025-01-24 for older models.

Everything else on the list above is GA, and the intuition that "the scary one must be the beta one" leads you straight to code execution, which is wrong. Release status tracks how settled the interface is, not how dangerous the tool sounds.

The pricing shape that changes designs

Look at search and fetch together, because the gap between them is a design instruction.

Search costs money per invocation. Fetch does not. Which means: discovery is billed, retrieval is not. An agent that already knows the URL and pulls it is free at the tool layer. An agent that searches the open web on every turn because it never bothered to remember anything has a line item.

That is not a micro-optimisation, and the reason is the loop. Everything inside an agent loop gets multiplied by turn count, and turn count is the number nobody bounds until the invoice arrives. A habit that costs a fraction of a penny per turn is invisible in a demo and a budget line in production. Caching a URL is not premature optimisation here; it is the difference between two pricing models.

The versioned type strings

Every built-in tool is referenced by a dated type string, and those dates move. You will see this concretely in the next lesson, where code_execution_20250825, code_execution_20260120 and code_execution_20260521 all exist, all listed GA, and the documentation is not consistent about which one a given feature wants.

Do not try to memorise the current date for every tool. Memorise that the dates exist, that several versions can be live simultaneously, and that this is a lookup on exam day. Knowing which facts are lookups is worth marks on a timed paper, because it stops you burning three minutes trying to recall something you were never going to recall.

Practice

Try it yourself

Quiz

The odd one out

Six of these are GA. One is not, and it is the one people assume is furthest along because it demos the best.

Recall

Search and fetch, priced

Two web tools, two very different bills. Exact numbers, because the exam quotes exact numbers.

What does web search cost, and what does web fetch cost?

Quiz

Price the habit

Your agent runs 20 turns per session and calls web search on 8 of them. You run 500 sessions. What does the search tool cost you, tokens aside?

Quiz

Two headers, two model generations

This is the kind of fact that is pure lookup in real life and pure recall on an exam.

Quiz

Two minutes, docs open, which version

You need the current type string for the code execution tool, you know several versions exist, and you have not memorised any of them.

Check

Audit your own agent

Take the research assistant from the capstone brief, or any agent you have actually built, and list every built-in tool it would use.

You should see

For each tool: is it GA or beta, does it need a beta header, and what does a thousand invocations cost you. If you cannot fill the cost column for web search without looking it up, that is the card to review tonight.