What Are AI Agents? How They Work and When to Use Them

1. Send the task, conversation and tool definitions to the model.
2. The model replies with a tool call (or a final answer).
3. Your code runs the tool and sends back the result.
4. Repeat until the model answers, or a step limit stops it.

The model never runs anything itself. Your code stays in control.

Anthropic's widely cited Building Effective Agents (2024) separates workflows, where code defines the steps (prompt chaining, routing, parallelization, orchestrator–workers, evaluator–optimizer), from agents, where the model directs its own steps. The advice: start with the simplest option that works. Many problems don't need an agent.

Agents need memory you build yourself: short-term (the conversation) and long-term (stored facts retrieved by relevance). Multi-agent systems split work across a lead agent and subagents. They're powerful for broad research but expensive: Anthropic reported its multi-agent research system used about 15× the tokens of a normal chat.

• Hard limits on steps, time and spend.
• Tool errors returned as text so the model can recover.
• Human approval for payments, deletes, emails and other irreversible actions.
• Least-privilege credentials and allowlists.
• Full tracing of every step, plus an eval set of realistic tasks.