What it is

An agent isn't a predefined path. It plans and operates independently, gaining ground truth from the environment at each step (tool results, code execution), and can pause for human feedback at checkpoints or when blocked. Anthropic's reassurance: implementation is often straightforward.

They are typically just LLMs using tools based on environmental feedback in a loop. — Anthropic

The agent loop (and ReAct)

goal ─▶ ┌──────────────────────────────────────────┐
        │  plan ─▶ act (tool) ─▶ observe ─▶ reflect  │ repeat
        └──────────────────────────────────────────┘
          exit when: final-output tool · no tool call ·
                     error · max turns reached

  ReAct (Yao et al.):  Thought ─▶ Action ─▶ Observation ─▶ ...

OpenAI describes this 'run' as a while-loop that runs the LLM until an exit condition (a final-output tool, a response with no tool calls, an error, or a max-turns cap).

What's inside an agent

Lilian Weng's anatomy: an LLM 'brain' plus three faculties.

  • Planning — task decomposition (Chain-of-Thought, Tree-of-Thoughts) and self-reflection (ReAct, Reflexion).
  • Memory — short-term (the context window) and long-term (external vector stores for fast retrieval).
  • Tool use — calling external APIs/functions to act beyond the model's frozen weights.

Single-agent vs multi-agent

OpenAI's advice: maximize a single agent (one model + tools in a loop) first; it keeps complexity and evaluation manageable. Split into multiple agents only when prompts get full of conditionals or tools overlap and confuse the model.

  • Manager (agents as tools) — a central manager calls specialized agents via tool calls and keeps control.
  • Decentralized (handoffs) — peer agents transfer control to one another based on specialization.
  • OpenAI: in the manager pattern edges are tool calls; in the decentralized pattern edges are handoffs.

When to use it (and the costs)

For open-ended problems where it's difficult or impossible to predict the required number of steps, and where you can't hardcode a fixed path. — Anthropic
  • Highest autonomy = highest latency, cost, and least predictability — use deliberately.
  • Needs guardrails, transparency (show the planning steps), and a well-designed agent-computer interface (ACI).
  • Weng's limits: finite context, hard long-horizon planning, and unreliable natural-language tool interfaces.

Concrete examples

  • Coding agents resolving real GitHub issues in SWE-bench Verified from a pull-request description alone.
  • Anthropic's 'computer use' reference implementation, where Claude operates a computer autonomously.
  • Verifiable domains (code with tests) suit agents because results feed back as ground truth.