Corpus Agentis
The field book to agent ecosystems
The field book to agent ecosystems
Stack · Agent tools

The agent toolbox

A model on its own only writes text. An agent is a model given tools: the capabilities it invokes to search, run code, drive a computer, read files and remember. Here is the common core, capability by capability, sourced to each provider's docs.

13 built-in tool types every serious agent platform exposes, grouped by what they do. 7 are near-universal (OpenAI, Anthropic and Google all ship them); the rest are emerging, shipped by one or two providers so far. Each is the capability itself, not a vendor product: the identifiers show how OpenAI, Anthropic and Google name it, and every row links to a provider's official tool documentation.

Information & retrieval · 3
Web search
● near-universal

Run live web queries and ground the answer in current sources with citations, rather than relying on the model's training cut-off.

Built-in as OpenAI web_search · Anthropic web_search_20250305 · Google google_search (grounding)
Web fetch
● emerging

Retrieve the full content of a specific URL the agent already holds, as opposed to searching the open web for one.

Built-in as Anthropic web_fetch_20250910 · Google url_context (OpenAI folds this into web_search)
File search / retrieval
● near-universal

Retrieve the relevant chunks from a developer's uploaded documents or vector store into the model's context (retrieval-augmented generation).

Built-in as OpenAI file_search · Google file_search (Anthropic via the Files API + code execution)
Execution · 3
Code execution
● near-universal

Write and run code, typically Python, in a hosted sandbox for computation, data analysis and mathematics.

Built-in as OpenAI code_interpreter · Anthropic code_execution_20250825 · Google code_execution
Shell / bash
● emerging

Execute arbitrary shell commands in a runtime environment, local or hosted, giving the agent a terminal.

Built-in as Anthropic bash_20250124 · OpenAI local_shell (Google folds this into code_execution)
Image generation
● emerging

Generate or edit images from a text prompt as a callable step inside the agent loop.

Built-in as OpenAI image_generation (Google treats image output as a model capability; Anthropic has none)
File I/O · 1
File read / write
● near-universal

View, create and edit files on disk via structured commands such as view, create, str_replace and insert.

Built-in as Anthropic text_editor_20250728 · OpenAI apply_patch
Control · 1
Computer use
● near-universal

Drive a graphical interface by taking screenshots and issuing mouse and keyboard actions to control a computer or browser.

Built-in as OpenAI computer_use · Anthropic computer_20250124 · Google computer_use
Memory · 1
Memory
● emerging

Persist and recall information across turns and sessions through a dedicated memory store the agent can read and write.

Built-in as Anthropic memory_20250818 (no API-level built-in from OpenAI or Google yet)
Integration & extensibility · 4
MCP connector
● near-universal

Connect the model to third-party tools and data exposed by a remote Model Context Protocol server, without hand-wiring each integration.

Built-in as OpenAI mcp · Anthropic mcp_toolset · Google via the Gemini SDK
Function calling
● near-universal

Call developer-defined functions with structured JSON arguments; the universal substrate every other built-in tool is built on.

Built-in as OpenAI function · Anthropic tools · Google function_declarations
Data / SQL access
● emerging

Query structured databases and warehouses; delivered through MCP database connectors or by running SQL inside code execution rather than a first-party built-in.

Built-in as via MCP connectors or code execution (no first-party built-in from the three)
Tool search
● emerging

Dynamically discover and lazily load the relevant tool definitions at runtime, saving context when many tools are available.

Built-in as Anthropic tool_search_tool_regex_20251119 / _bm25_20251119 · OpenAI tool_search (gpt-5.4+)
Field notes

The built-in tools every agent calls: search, code, computer use, files, memory and MCP

Thirteen built-in capabilities, the things an agent can actually do once it has a model: search and fetch the web, run code, use a shell, read and write files, drive a computer, retrieve from files, hold memory, call MCP connectors, generate images, call functions, query data and search its own tools.

Show more

These are the primitives every vendor now ships in some form, which is why the interesting question is no longer whether an agent has them but what it is permitted to do with them. Each one is also an entry point: web fetch and file read are how untrusted content reaches the model in the first place.

From the corpus, curated by Brandon Chaplin
Common questions
What built-in tools do AI agents have?

A common core recurs across providers: web search, web fetch, code execution, a shell, file read and write, computer use, retrieval, memory, image generation and MCP connectors. Each vendor names them differently but the capabilities line up. Function calling is the mechanism underneath all of them.

What is the difference between a built-in tool and function calling?

Function calling is the mechanism: the model emits a structured request naming a tool and its arguments. A built-in tool is a capability the provider ships and often runs on its own servers, such as web search or code execution, exposed through that same mechanism. Custom tools you write use the same substrate. (OpenAI)

Which tools do all the major AI platforms provide?

Web search, code execution, computer use, file search and function calling are shipped by OpenAI, Anthropic and Google alike. MCP connectors are close behind. Memory, standalone web fetch and tool search are still uneven, offered by only one or two providers.

How do agents get tools their model provider does not ship?

Through MCP. An agent connects to a server that exposes third-party tools, so database access or a CRM integration works without the model provider building each one. Where no server exists, running code inside the sandbox fills the gap. (Model Context Protocol)

Is an agent tool the same as an agent framework?

No. Tools are the capabilities an agent invokes. A framework is the library you build the agent on, deciding when and in what order those tools get called. LangGraph and the OpenAI Agents SDK are frameworks; web search is a tool.

Next in the learning path