Egor Urvanov

Tool and tool calling

Agents and tools · tool call, tool use, function calling

A tool is an action the agent can call: read a file, find a ticket in Jira, run a database query. The set of tools defines what the agent can do and what it can reach. Tool calling (function calling) works like this: the model returns a function name and arguments, for example search_jira with the query “login”; the harness runs the call and passes the result back to the model.

JiraLLMthe model sees descriptions of:read_file · search_jira · run_sqlread_fileproject filessearch_jiraJirarun_sqldatabasecall: search_jira("login")Jira searchresult: 3 ticketssend_email: not grantedemail✕● model, people, systems, data  ● agent and harness  ● risk, stop, boundaryThe model sees only tool descriptions, the harness performs the action

How a tool call works

  • 1. Description · along with the task, the harness sends the model a list of tools. For each: a name (search_jira), what it does (“searches Jira tickets”) and parameters (query, a search string).
  • 2. The model’s call · the model returns a call as JSON: {"name": "search_jira", "arguments": {"query": "login"}}.
  • 3. Execution · the harness checks the call against permissions and queries Jira. Writes and deletions are shown to a person for approval first.
  • 4. The result · the harness adds Jira’s answer (“3 tickets: …”) to the context and runs the model again. The model calls the next tool or writes the answer.
  • Which models support it · models from OpenAI (GPT), Anthropic (Claude), DeepSeek and other current LLMs; local ones through Ollama, if the model supports tools.
  • Tool calling vs MCP · tool calling is the call format from steps 2–4. MCP is a protocol through which an outside system (Jira, a 1C database) gives the agent its own list of tools and runs their calls.

Source: Anthropic: Tool use docs

Where it appears · 13

More 8

Related terms

Whole glossary

© 2026 Egor Urvanov