What is agentic AI?
Regular Claude use is reactive — you ask, Claude answers. Agentic use is different: Claude gets a goal and makes its own decisions about which steps are needed to reach it. Claude can call tools, evaluate results and adjust course along the way.
💬
Reactive: "What is the weather in Oslo?" → Claude answers from training data.
🤖
Agentic: "Find the weather in Oslo and send me an SMS if it rains." → Claude calls weather API, evaluates the response, calls SMS API if necessary.
How tool use works
💡 Tool use flow:
1. You define tools with name, description and parameters
2. Claude decides if a tool is necessary
3. Claude returns a tool_use call with parameters
4. You execute the call and return the result
5. Claude uses the result in its final answer
Security for agentic use
⚠️ Important security rules:
— Never let the agent perform irreversible actions without confirmation
— Limit which systems the agent has access to
— Log all tool calls for auditability
— Set maximum limits for number of iterations
— Always test in a sandbox environment before production
"An agent that can do everything is dangerous. An agent with precise tools and clear boundaries is powerful."
— On responsible agentic design