Skip to main content

Understanding Tools

Tools are external capabilities that an agent (an LLM-driven system) can call to perform actions beyond text generation. They give an LLM practical power: access to up-to-date data, precise computation, side effects (APIs, databases, file systems), and longer-term state.

How tools empower LLMs​

  • Extend knowledge: provide current facts (search, web browser, knowledge bases).
  • Perform exact work: calculators, spreadsheets, code runners, and validators remove guesswork.
  • Take actions: APIs, databases, and OS tools let agents effect changes (create tickets, send emails, deploy).
  • Maintain context/state: tools can record or query durable state outside the model context window.

Roles of tools in an agent​

  • Observers: retrieve information (search, sensors, logs).
  • Actors: perform operations with side effects (API calls, job scheduling).
  • Translators/Adapters: convert LLM output to structured inputs or protocol messages.
  • Validators: check correctness and enforce constraints (type checkers, linters, test runners).

Benefits​

  • Grounding: reduces hallucination by sourcing answers from authoritative tools.
  • Accuracy: deterministic tools (math, exact lookup) improve precision.
  • Scalability: offload heavy work to specialized services.
  • Safety: isolate risky operations and enforce checks before performing side effects.

Example tool manifest (simple):

{
  "name": "search",
  "description": "Web search returning top 5 results",
  "inputSchema": { "query": "string" },
  "outputSchema": { "results": "array" }
}

Tools transform an LLM from a text predictor into an actionable, reliable system component. Design them to be explicit, safe, and composable.