unknown

AI agents extend LLMs by placing them in a Think → Act → Observe loop with access to tools (APIs, databases, code execution), enabling them to complete multi-step tasks rather than just generate text. The three core components are the LLM brain, the tools it can call, and memory/state tracking. Real-world examples include Klarna's customer service agent (which handled 2.3M conversations but required human rehiring for complex cases) and coding agents like Cursor and Claude Code. Key failure modes include tool call hallucinations, compounding errors across steps, and organizational gaps around auth, audit trails, and compliance. A minimal Python implementation using LangChain's ReAct agent is shown. The pattern that works in production: constrained domains, clear tool definitions, and humans in the loop for edge cases.

What is an AI Agent?

Paolo Perrone

### Everyone's building one. Almost nobody can define one.

<p>### Everyone's building one. Almost nobody can define one.</p>

<p>Curious how teams are deciding the right balance between autonomy and human-in-the-loop for AI agents in production. What guardrails have actually worked for you when agents start chaining multiple tools?</p>
