Agent Tool Categories
Agent tools fall into knowledge augmentation, capability extension, and write actions, with risk rising sharply for tools that mutate the environment.
Tools in an agent's tool inventory group into three practical categories:
Knowledge augmentation expands what the agent knows: retrievers, SQL executors, people search, inventory APIs, email readers, and web browsing. These construct context from private or public data. Web browsing fights training-cutoff staleness but exposes the agent to unreliable or harmful content, so API choice matters.
Capability extension patches model weaknesses more cheaply than training them away: calculators, calendars, converters, translators, and especially code interpreters that return execution results and errors. Tools can also make a unimodal model effectively multimodal (image generation, OCR, transcription, charting via code). Scaling alone does not fix up-to-date knowledge, precise math, low-resource languages, or time awareness — Toolformer-style tool use is one way to teach the model when those APIs help without task-specific wiring. Schick et al.’s canonical five: QA (Atlas), calculator (+−*/), BM25 Wikipedia search, NLLB→English MT, and a zero-arg calendar — all text I/O with a few demos.
Write actions change data sources rather than only reading them: mutating tables, sending email, initiating transfers. They unlock full workflow automation, but blast radius scales with autonomy — models are probabilistic and often wrong, so do not book travel or merge PRs merely because the user mentioned wanting to. Treat write access like production credentials for a junior operator — trust must be earned through sandboxing, human approval gates, and security measures, not assumed because the model is capable.
Sources
- AI Engineering (Chip Huyen), ch. 6
- Prompt Engineering for LLMs (Berryman), ch. 4
- Toolformer (Schick et al.), pp. 1–17