Agent loop

An agent loop is the core execution pattern of an AI agent: a repeating cycle of thinking, acting, and observing that continues until the agent reaches a stopping condition.

The loop:

1. Think. Given the current state + goal, the LLM decides the next action.
2. Act. The agent calls a tool — an API, a database query, a code execution, etc.
3. Observe. The tool returns a result that becomes new context.
4. Repeat. Loop until the goal is reached, the agent decides to escalate, or a step / time / cost budget is exceeded.

Why the loop matters: it's what separates an "AI feature" (one LLM call, one response) from a real agent (multiple coordinated steps to accomplish a goal).

Production agent loops include guardrails:

• Step limits. Maximum N tool calls per task (prevents infinite loops).
• Cost limits. Maximum spend per task.
• Confidence checks. If the agent isn't confident, escalate rather than continue.
• Tool sandboxing. Some tools (e.g., sending external email) require human approval mid-loop.
• State persistence. For long-running agents, state is saved so the loop can resume across sessions.

The MCP (Model Context Protocol) standard from Anthropic is the emerging way to standardize how agents discover + interact with tools across the loop.