Knowledge
Knowledge HubConcepts and strategic researchGlossaryDefinitions across six domainsFrameworksStructured strategy methodsTopicsConnected subject areasLibrary
PlaybooksPractical strategy guidesBooksBooks and long-form researchOperator VaultMember resources and toolkits
Explore Discover, the Knowledge Hub, playbooks, and the Operator Vault.
Glossary
Tool calling
Tool calling is the pattern in which a model emits a structured request to use an external function or API (name, arguments, and related metadata), while the host application or runtime validates permissions, executes the tool, and returns the result to the model or workflow. Closely related phrases such as “tool use” describe the broader capability of models acting through tools; they are not always exact synonyms for this structured call–execute–return loop.
Why it matters
Without tools, generative models are limited to text in and text out. Tool calling connects generation to systems of record — search, CRM, calendars, calculators, deployment APIs — which is why agents and many enterprise assistants depend on it. Accountability follows the host: the model proposes; the execution environment decides whether the call is allowed and what side effects occur.
Operators should treat tool schemas as a security boundary: argument validation, least-privilege credentials, logging, and human approval for irreversible actions matter as much as prompt quality. Confusing “the model invoked the API” with “the host executed an authorized call” blurs incident ownership.
Example (enterprise): A scheduling assistant’s model returns a structured `create_calendar_event` request. The host checks the user’s OAuth scopes, rejects disallowed calendars, creates the event only if policy passes, then feeds success or error back into the next model turn.
Mechanisms
- Schema declaration — tools are described so the model can select names and arguments in a structured format.
- Model emission — the model produces a tool-invocation request rather than (or in addition to) final user-facing text.
- Host execution — the application validates, authorizes, runs the tool, and supplies the observation/result.
- Loop continuation — results return to the model or orchestrator for the next step or final answer.
Boundaries
- Tool calling ≠ unconstrained code execution; production systems constrain which tools exist and who may run them.
- Tool calling ≠ RAG by itself, though retrieval can be implemented as a tool.
- Tool calling ≠ “tool use” as a guaranteed exact synonym — tool use can name broader or informal patterns.
- Emitting a tool request is not the same as successful, authorized execution.
Related topics
None linked yet.
Related concepts & frameworks
Further research
No related resources yet.