Hunter Brennick AI Systems & Advisory ↗
AI Systems Orchestration
Part II · Patterns And Practice/Chapter 04
04

Agent And Workflow Patterns

The pattern language: workflow vs agent, prompt chaining, routing, parallelization, orchestrator workers, and evaluator optimizer loops.

4 min read2 figures
The gist
  • Workflows buy control. Agents buy flexibility. Hybrids buy both.
  • The first question is not "how many agents?" It is "where should control live?"
  • Multi agent helps only when specialization improves tools, context isolation, evals, or parallelism.
THE PATTERN LANGUAGE PROMPT CHAINING model step model step model step each step feeds the next · easy to inspect between steps ROUTING router specialist A specialist B specialist C classify first · send each input to the handler built for it PARALLELIZATION split part one part two part three merge independent parts run at once · latency of the slowest part ORCHESTRATOR · WORKERS orchestratorplans and delegates worker worker worker work is not known up front · the orchestrator decides and synthesizes EVALUATOR · OPTIMIZER generatordrafts the answer evaluatorscores against criteria candidate feedback · try again acceptedpasses the bar workflows first · add autonomy only where it earns its complexity
FIG 04.1The five canonical patterns: chaining, routing, parallelization, orchestrator workers, and evaluator optimizer.

Module 4: Agent And Workflow Patterns

This is where the common vocabulary lands. Chapter 0b gave the one-sentence version: workflows pick the steps, agents pick their own. This chapter expands that into a pattern language.

Read
Anthropic: Building effective agentsanthropic.com↗OpenAI practical guide to building agentsopenai.com↗Microsoft Agent Framework overviewlearn.microsoft.com↗

Learn the patterns that give this vocabulary its shape:

  • prompt chaining
  • routing
  • parallelization
  • orchestrator-workers
  • evaluator-optimizer
  • a manager agent with specialists as tools
  • handoff between specialist agents
  • an autonomous loop with stop conditions

The key distinction:

WHERE SHOULD CONTROL LIVE? more control more flexibility Workflow predefined process code decides the next step reliable · auditable · testable Agent model decides the next step handles open ended decisions flexible · adaptive · harder to bound Hybrid system code owns the shape · model owns the judgment steps are known decisions are open ended
FIG 04.2The key distinction: workflows buy control while agents buy flexibility and hybrids buy both.

Build by picking five tasks you regularly give agents and labeling each one:

  • deterministic workflow
  • single agent
  • manager plus specialists
  • graph runtime
  • human in the loop

For each task say what would make it unsafe or too expensive.

Multi-agent is not automatically better. It helps only when specialization improves:

  • tool clarity
  • context isolation
  • evaluation
  • parallelism

The first question is not how many agents to use. It is where control should live.

Next chapterChapter 05 · Context, Tools, And Protocols Where usefulness lives or dies: knowledge sources, context engineering, memory, tool shape, MCP, A2A, Skills, and AGENTS.md.