10Agent Design Patterns & Multi-Agent Systems
Prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer and multi-agent teams — when to use each, with business examples.
What you'll learn
- Recognize the six core agentic patterns
- Pick the simplest pattern that solves a problem
- Understand when multi-agent systems help (and when they hurt)
- Plan observability and cost control
Patterns, from simplest to most autonomous
Most production AI systems are built from a small set of reusable patterns. Knowing them lets you design systems on paper before writing code.
1 · Prompt chaining
Fixed sequence of steps, each consuming the last. Use for: content production, report generation, translation + localization.
2 · Routing
Classify first, then send to a specialized handler. Also great for cost: simple questions to a small model, hard ones to a big model. Use for: support desks, inbox triage, lead qualification.
3 · Parallelization
Run several calls at once — either sectioning (different sub-tasks in parallel) or voting (same task several times, compare answers). Use for: reviewing a contract for five risk types simultaneously; content moderation where you want high confidence.
4 · Orchestrator–workers
A lead model breaks the job into sub-tasks it can't predict in advance, delegates them, then combines results. Use for: market research across many sources, large codebase changes.
5 · Evaluator–optimizer
One model generates, another critiques against criteria, loop until it passes. Use for: proposals, ad copy, translations — anything with clear quality criteria.
6 · Autonomous agent
The full loop from Lesson 06: the model plans and uses tools freely within guardrails. Use for: open-ended tasks where steps depend on discoveries.
| Pattern | Predictability | Cost | Pick it when… |
|---|---|---|---|
| Prompt chaining | High | Low | Steps are known and sequential |
| Routing | High | Low | Inputs fall into distinct categories |
| Parallelization | High | Medium | Sub-tasks are independent, or you want consensus |
| Orchestrator–workers | Medium | Med–High | Sub-tasks can't be known in advance |
| Evaluator–optimizer | Medium | Medium | Quality criteria are clear and iteration helps |
| Autonomous agent | Lower | Highest | Open-ended goals with tools and verification |
Multi-agent systems
A multi-agent system uses several agents with different roles, instructions or tools — e.g. a researcher, a writer and a reviewer, or a planner with specialized executors. Frameworks like CrewAI, AutoGen and LangGraph make these easy to set up.
When they help
- Work splits naturally into parallel, independent parts
- Each part benefits from focused instructions and tools
- The context would be too large for one agent
- You want a separate reviewer that didn't produce the work
When they hurt
- Tasks are tightly coupled and need shared context
- Coordination overhead costs more than it saves
- Errors compound between agents with no verification
- You can't trace which agent did what
See a simulated researcher → writer → reviewer team in the AI Lab.
Running agents responsibly: observability & cost
Trace every run
Log each model call, tool call, arguments, result, tokens and time. Tools like LangSmith, Langfuse or simple JSON logs all work.
Budget per run
Cap steps, tokens and money. Alert when a run approaches the limit.
Use the right model per step
Small/cheap (or local) models for classification and extraction; strong models for planning and hard reasoning.
Cache what repeats
Long, stable system prompts and documents can often be cached by the provider to cut cost and latency.
Review failures weekly
Read failed traces, add the case to your test set, fix the prompt or tool.
Key takeaways
- Six patterns cover most systems: chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, autonomous agent.
- Choose the most predictable pattern that works; add autonomy only when needed.
- Multi-agent helps with parallel, separable work — but adds coordination cost and failure points.
- Trace, budget, route by difficulty and review failures to keep agents reliable and affordable.
Knowledge check
0 / 3Q1Support tickets fall into billing, technical and sales. Which pattern fits first?
Routing classifies inputs and sends them to specialized handlers.
Q2Generate → critique → revise until it meets criteria is called…
An evaluator scores the output and the optimizer improves it in a loop.
Q3A key risk of multi-agent systems is…
More agents = more handoffs, cost and places for errors to spread.