What Are AI Agents?
The agent loop, autonomy levels, and when to use agents vs simple prompts
An AI agent is a model that can take actions and observe results — repeatedly — until it achieves a goal. The defining characteristic is the loop: perceive state → reason → act → observe new state → repeat. This is fundamentally different from a single prompt-response cycle. Understanding when to use agents and when not to is the first skill of agent engineering.
The agent loop
- Perceive — The agent receives context: the task, current state, available tools, and any results from previous steps.
- Reason — The model decides what to do next: which tool to call, what to compute, whether the task is complete. This reasoning happens in the model's "thinking".
- Act — The model calls a tool or produces output. Tools might be web search, code execution, database queries, API calls, or spawning sub-agents.
- Observe — The tool result is appended to context. The loop repeats with this new information. Loops continue until the agent signals completion or hits a step limit.
Autonomy levels
- Level 0: Single prompt — One input → one output. No tool use, no loops. Appropriate for well-defined transformation tasks.
- Level 1: Tool use (one round) — The model calls one or more tools, gets results, produces final output. Most AI features are here.
- Level 2: Agentic loop (multi-step) — The model runs multiple tool calls in sequence, reasoning between steps. Appropriate for research, code generation with testing, and multi-step workflows.
- Level 3: Multi-agent (hierarchical) — An orchestrator agent delegates to specialist sub-agents. Each sub-agent has its own tools and context. Appropriate for complex, long-horizon tasks.
When NOT to use agents
- When a single prompt is sufficient — If the task has a clear input and a clear output with no need for multiple steps or tool calls, an agent adds complexity without benefit.
- When latency is critical — Each loop iteration adds latency. A 5-step agent with 1 second per step is 5+ seconds. If the user needs an answer in under 2 seconds, use a single prompt.
- When determinism is required — Agents make autonomous decisions. If every step must be predictable and auditable, use structured orchestration code rather than agent reasoning.
- When cost is constrained — Multi-step agents consume tokens at every step. A 10-step agent costs 10x more in token terms than a single prompt. Calculate before committing.
Start simple — add agent loops only when needed
The best agent systems start as simple prompt chains and graduate to loops only when the problem genuinely requires autonomous decision-making. Premature agentification is the main source of unpredictable, expensive, hard-to-debug AI systems.
Try this
Take a task you are considering building as an agent. Map it as: what is the input, what are the steps, what decisions require AI reasoning vs deterministic code? If fewer than 3 steps require genuine AI reasoning, you probably do not need an agent loop — a prompt chain will do.