AI Agent Tools Examples in 2026

Ten proven agent tool implementations with key design points: web search, PR review, SQL, calendar, email, document conversion, image generation, browser control, vector retrieval, and sandboxed code execution.

💡 What You Will Learn

Ten proven agent tool implementations with key design points: web search, PR review, SQL, calendar, email, document conversion, image generation, browser control, vector retrieval, and sandboxed code

An agent's ceiling is set by its tools. Ten battle-tested implementations:

1. Web search — Tavily (purpose-built for agents; free tier per official site) or self-hosted search API + content fetching. Return title/summary/URL only, never full HTML.
2. GitHub PR review (see PR-Agent, ~12k stars) — pull the diff, review file by file, grade comments as must-fix/suggest/ignore.
3. Read-only SQL — read-only account, parameterized queries, forced LIMIT, table whitelist. Never let the agent write.
4. Calendar booking (MCP Calendar) — check availability first, always require user confirmation before booking.
5. Email triage — keep 3-5 labels, fixed summary templates, don't send spam to the model.
6. Document conversion (markitdown, ~30k stars, Microsoft) — PDF/Word/PPT to Markdown before feeding RAG; many "can't read PDF" problems are format problems.
7. Image generation via ComfyUI API — templated prompts, fixed negative prompt, basic output validation.
8. Browser automation (browser-use, ~100k stars) — small steps, screenshot checks per step, timeouts; great for forms, scraping, testing.
9. Vector retrieval — chunk 256-512 chars, top-k 3-5, filter relevance before injecting into context.
10. Sandboxed code execution (E2B, ~13k stars) — isolated cloud sandbox, destroyed after run; never execute agent-generated code on the host.

Common combos: search + vector retrieval for research; browser + document conversion for page analysis; SQL + sandbox for data analysis.

Three rules: keep tools small with clear docstrings (when to use, parameter format); return error messages to the model instead of empty results (the agent can retry or switch tools); always sandbox execution tools.

FAQ: start with 2-3 tools, add more after it works. Truncate and summarize long tool output. On tool errors, return the raw error text — the model adapts its strategy.

❓ FAQ

Which example is easiest to start with?

The SQL tool and the vector search tool - both are under 10 lines.

Are tools safe by default?

No. Sandbox any code execution and add allow-lists to shell tools.