# Agent patterns: from fixed workflows to autonomous agents

Most reliable agent systems are not fully autonomous. They combine a few well-known patterns, from fixed workflows where the developer sets every step to open loops where the model chooses.

## Overview

Agent patterns sit on a line. At one end the developer fixes the order of steps and the model fills each one in. At the other end the model decides every step. Moving right buys flexibility and costs predictability, latency and money, so the rule of thumb is to use the simplest pattern that solves the task.

## The patterns, from fixed to free

| Pattern | Who decides the steps | Good for | Main risk |
|---|---|---|---|
| Prompt chaining | Developer | Fixed multi-step tasks such as draft → check → format | Errors carry forward |
| Routing | Model picks one branch | Sending each request to the right handler or model | Wrong route |
| Parallel calls | Developer | Independent sub-tasks, voting for reliability | Cost |
| [ReAct loop](/tools/react-loop) | Model, one step at a time | Open questions that need tools | Loops that never end |
| [Plan and execute](/tools/plan-and-execute) | Model plans, then follows | Long tasks with clear stages | Plans that go stale |
| [Reflection](/tools/reflection) | Model critiques its own output | Writing, code, answers that can be checked | Extra cost per turn |
| [Orchestrator and workers](/tools/orchestrator-workers) | Supervisor model | Tasks that split into specialist work | Lost context at hand-offs |
| Swarm | Agents hand off to each other | Conversations that move between domains | Hard to trace |

## How to read the table

**Workflows first.** If you can write the steps down, write them down: prompt chaining and routing are cheap, testable and fast. **Add a loop where the path is unknown.** A ReAct loop lets the model search, read and decide what to do next. **Plan when the task is long.** A plan gives the run a shape and lets you check progress against it; re-plan when a step fails. **Reflect where quality can be judged.** A second pass that checks the answer against criteria catches many errors, at the price of another model call. **Split into agents only when the work really splits.** Several agents add flexibility, but every hand-off can drop information; a supervisor that keeps the shared state reduces that loss. See [sub-agents](/tools/sub-agents) and [multi-agent systems](/tools/multi-agent-systems).

## Stopping conditions

Every loop needs three exits: the goal is met, a budget runs out (steps, time, tokens or money), or the agent needs a person. Agents without explicit budgets are the main cause of runaway cost. See [cost and latency budgets](/tools/cost-and-latency-budgets) and [human in the loop](/tools/human-in-the-loop).

## Learn it hands-on

The [agentic patterns simulation](/ai-simulations/agentic-patterns) runs several patterns on the same task, the [multi-agent supervisor simulation](/ai-simulations/multi-agent-supervisor) shows routing between specialists, and [one agent run, traced](/ai-simulations/agent-run) shows a ReAct loop step by step. [Tree of thoughts](/tools/tree-of-thoughts) extends planning with search over alternatives.

## Frequently asked questions

### Do I need a framework for these patterns?

No. Chaining, routing and a basic loop are a few dozen lines of code. Frameworks help when you need durable state, approval steps and tracing.

### Is a multi-agent system better than one agent?

Only when the task divides cleanly into specialist parts. For most tasks one agent with good tools and a clear plan is cheaper and easier to debug.

## Summary

Start with fixed workflows, add a loop where the path is unknown, plan for long tasks, reflect where quality can be checked, and split into several agents only when the work truly splits.

## Back to the topic

This page is part of **[Agentic AI](/topics/agentic-ai)**, which maps every concept, tool, simulation and article on the subject.
