Tools and function calling · AI Tools Field Guide

The model writes a request; the harness does the work; the result comes back as text.

The round trip

A model produces text. That is all it does. Every ability beyond text, from searching to reading a file to running a calculation, is a tool: a function the harness owns and offers to the model. The mechanism is the same everywhere.

The harness lists the tools in the request, each with a name, a one-paragraph description, and a schema for its inputs. This list is text, and it sits in the preamble layer ahead of your message.
The model decides to call one. Instead of prose it emits a structured request: the tool's name and the input values. It is still predicting tokens; it has learned that this shape, here, is the likely next thing.
The harness validates and runs it. The model never touches your files or the network. The harness checks the inputs against the schema, executes the function, and catches the errors.
The result comes back as a message, appended to the context like any other turn. Large results are cut to a head and a tail before they go in.
The model continues with the result in front of it, and may call again.

That is function calling. Step one and step two look like this, trimmed to the parts that matter:

{"name": "search_policy",
 "description": "Search the command's policy library. Use for questions about instructions, notices and SOPs.",
 "parameters": {"query": {"type": "string"}, "max_results": {"type": "integer"}}}

{"tool": "search_policy", "input": {"query": "leave request routing", "max_results": 5}}

Two things follow. The description is a prompt: the model chooses a tool by reading it, so a vague description gets a vague choice. And the inputs are the model's guess: a wrong date or a made-up file name in a call is the same hallucination you already know from prose, dressed as a parameter.

Where calls go wrong

Failure	What you see	What fixes it
Wrong tool chosen	It searched when it should have read the attached file	Tighter descriptions; fewer tools
Invented input	A file name or date that does not exist	Make the tool reject bad inputs; make the model confirm
Truncated result	It answers from the first page of a long document	Ask it to search within the result, or split the file
Result read as instruction	A web page or document steers the next call	Narrow tools, review the transcript; see prompt injection
No stop condition	It calls, reads, calls again	A budget and a test for done; see Agents and automation
No call at all	"According to the instruction…" with no tool call in the transcript	It answered from training. Ask it to look, then check the citation.

The last row is the one to memorize. If there is no tool call in the transcript, nothing was searched and nothing was read.

Connectors and MCP

A connector is a tool that reaches something outside the platform: a calendar, a drive, a ticket system. Most connectors now speak one standard, the Model Context Protocol (MCP): a server describes its tools in the schema form above and any harness that speaks the protocol can offer them to any model. Two consequences. Every connector you switch on adds its tool descriptions to every request, so the tool list grows and the window shrinks. And every connector is a new way for the model to act on the world, so the narrow-tools rule from the table is the one to work by.

Which platform brings what, as of September 2026

Each vendor's product changes what it has switched on over time. The table holds what the vendors have said publicly; the product itself is the authority on any given day.

Platform	Reported abilities	Suited to
ChatGPT	Chat, files, projects and custom GPTs, with further features sequenced over time; web search with citations	Document-heavy work: planning, policy, logistics, admin, forms, drafting
Grok	Reasoning modes (Auto, Fast, Expert), workspaces, persistent projects, reusable playbooks, Deep Search (GovConWire)	Research and analysis tasks with several steps
Gemini	Gems: no- and low-code agents for multi-step tasks, shareable to named users or a whole organization (Google help)	Reusable helpers for read-aheads, action items, checklists, document review
Claude	Projects, artifacts, connectors, audit logs on managed plans	Document-heavy work, coding, and structured multi-step tasks. Ask your CIO shop about the plan your organization has. (Anthropic)

Read the table as "who has the tool for this job today", not as a ranking. A retrieval tool with citations is the thing to reach for when the answer must trace to a document; a reasoning mode when the job has steps; an agent builder when the job repeats and other people should be able to run it.

What to do with this

Open a transcript from last week and find every tool call. For each one, ask whether the tool was the right one and whether the inputs were real. Then take one job and run it once with a connector on and once with it off, and read the difference. Next: putting tools in a loop, Agents and automation.

Vendor facts checked 16 Sep 2026; re-verify before relying on them.