What is tool use? Querying and acting through external programs
Understand tool requests, execution environments, evidence, argument permissions, errors, and untrusted returned content through weather and file examples.
Understand the concept
Tool use lets AI applications obtain information or act through programs. A model often selects a tool and arguments, a runtime executes it, and results return to the model. Tools may be hosted or application-run. A claimed lookup is not execution evidence, and results need status and provenance checks rather than automatic trust or authority.
Model requests and tool execution are separate stages
For weather, a model needs an available query tool and location, followed by external data. Generated weather prose without execution is not a lookup. This is an example without live access or invented observations.
Anthropic distinguishes client and server tools, while OpenAI documents application functions and built-in tools. Establish execution ownership and return formats; otherwise a model-generated request may be mistaken for automatic local execution.
Constrain tool capabilities by identity and task scope
File tools should validate allowed paths rather than accept arbitrary directories. Queries need no deletion access. Descriptions guide selection but do not enforce authorization; arguments need type and business validation like other external input.
For side effects, consider timeouts and retries. A delayed write response does not establish failure. Reconcile through execution IDs or target queries, and return explicit errors rather than success-shaped defaults.
Inspect content and provenance after results return
Fetched pages and files may contain misleading instructions; retrieval does not elevate their authority. Preserve provenance and distinguish execution state from claims in text. A page saying success does not establish the application’s own completed write.
Start validation with known read-only data and check error paths and matching result IDs. No external tool experiment was run here. Real acceptance depends on target state or reproducible results; model summaries communicate rather than replace evidence.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.