Orchestration Patterns
A map of the main agent orchestration patterns — routing, chaining, parallelization, orchestrator-workers, and evaluator-optimizer — and which project on this site teaches each
Orchestration Patterns
TL;DR
"Orchestration" means coordinating several Large Language Model (LLM) calls or agents to solve one problem. There is a small, well-known set of patterns for doing it. This page maps each pattern to a project on this site so you can see it working — and, just as important, learn when to use which (and when to use none).
Start Simple
The most important rule in orchestration is also the easiest to forget: use the simplest design that works, and add structure only when a simpler one clearly can't cope. Most real-world failures come from over-engineering, not from using too small a model.
Escalate only when you have to
Workflows vs Agents
It helps to split orchestration into two families:
Two families of orchestration
Workflows
The path is decided in advance. You wrote the steps; the LLM just fills them in. Predictable, cheap, easy to trace. Routing and prompt chaining live here.
Agents
The path is decided at runtime. The system plans its own steps based on what it finds. More flexible, but harder to predict and debug. Orchestrator-workers and reflection loops live here.
The Five Patterns
A widely-used taxonomy (from Anthropic's Building Effective Agents) names five core patterns. Here is each one in plain English, when to reach for it, and where to see it on this site.
| Pattern | What it is | When to use it | See it here |
|---|---|---|---|
| Prompt chaining | Break a task into fixed, sequential steps; each step's output feeds the next | The task has clear, ordered sub-steps | Text Summarization (map-reduce / refine chains) |
| Routing | One classifier sends each request to the right model or path | Inputs fall into distinct categories | LLM Router, Adaptive RAG |
| Parallelization | Run sub-tasks at the same time — sectioning (split the work) or voting (many candidates, then combine/select) | Speed matters, or you want more confidence | Clinical Orchestrator (sectioning), Speculative RAG (parallel drafts → select) |
| Orchestrator-workers | A coordinator dynamically breaks the task down and delegates to workers | Sub-tasks aren't known until you read the input | Clinical Orchestrator, Multi-Agent System, Adverse Event Surveillance |
| Evaluator-optimizer | One agent produces, another critiques; loop until good enough | Quality improves with feedback and you can judge it | Clinical Orchestrator (critic loop), Self-RAG, Corrective RAG |
Real systems often combine patterns. The Clinical Orchestrator alone uses orchestrator-workers (the planner), parallelization (specialists), and evaluator-optimizer (the critic) together.
Two More Shapes Worth Knowing
Beyond the five, two coordination shapes show up constantly in practice:
- Supervisor / worker — a manager agent assigns and reviews specialist workers. See Multi-Agent System and Adverse Event Surveillance.
- Debate / adversarial — agents argue opposing positions and a judge decides, which surfaces disagreement instead of hiding it. See the Tumor Board Simulator and the other debate projects.
Know How They Fail
Adding agents adds failure modes. A 2025 study — the Multi-Agent System Failure Taxonomy (MAST) — found 14 common failures — weak verification, infinite loops, role drift, lost context — and concluded that most stem from design, not model quality. Before you ship any orchestration, read the failure-modes breakdown and the countermeasures in the Clinical Orchestrator project.
Where to Start
- New to agents? Begin with the Tool Calling Agent.
- Want the simplest orchestration first? Build the LLM Router.
- Ready for the full picture? Build the Live Clinical Orchestration Simulator.