Tools
Role prompts without tools are theater. Production crews earn trust by calling real systems and showing the observation.
A mechanic with a toolbox vs a mechanic who only describes engines. The tool is the wrench.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
Key Takeaways
- 1.An LLM alone returns text. An agent with tools can search, read files, hit APIs, and come back with observations. Built-in tools cover search, web, files. Custom tools are Python functions the loop may call.
- 2.The loop is: reason → choose tool → execute → observe → reason again. This is where a crew stops being conversational and starts being useful.
- 3.Attach a tool only to the agent that should have that permission. The docstring is what the model reads.
- 4.Return short observations — do not dump HTML. Fail closed on secrets and writes.
Learn elsewhere
- →Tool Calling — Phase 7
- →Least Privilege — Phase 20
Real Example
Scenario
Researcher: I need current information → search_web() → observation → reason → research complete.
What you would do
In CrewAI, apply Tools to this scenario: Researcher: I need current information → search_web() → observation → reason → research complete. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Tools make agents usefulDocstring is the schemaObserve then reasonLeast privilege per role
Cheat Sheet
Quick recap
quick ref- •Think-act-observe
- •Short observations
- •Split tools by role
- •Not every agent searches
Common Mistakes
- ✕Skipping evaluation for Tools before production
- ✕No logging or tracing around crewai tools steps
- ✕Ignoring cost and latency implications