Skip to content

Workflow Architecture Patterns

These patterns are the playbook’s one catalogue of how AI work is laid out, from a single augmented LLM call to agents that plan their own steps. Most of the names come from Anthropic’s Building effective agents; Handoff, Group Chat and Swarm come from the multi-agent guides by OpenAI, Microsoft and Google. Each page lists the other names vendors use.

The number of agents never changes the pattern. A prompt chain can be three LLM calls or three agents, and one page covers both.

The right pattern depends on what your workflow actually needs, not on how sophisticated you want it to be.

Two principles come before any pattern:

  1. Start with the simplest thing that works, and add complexity only when it measurably helps. Try a single Augmented LLM or a Prompt Chain first, and move up only when it falls short. Anthropic recommends “finding the simplest solution possible, and only increasing complexity when needed” (Building effective agents).
  2. One agent before several. Split the work only for a named reason: protecting context, parallel coverage, specialization, or a security boundary. Coordination has a cost: Anthropic reports that in its own data a single agent used about 4 times the tokens of a chat, and a multi-agent system about 15 times (How we built our multi-agent research system, June 13, 2025).

The tree asks the same question as the AI Workflow Design Matrix: what decides the next step?

What decides the next step?
│
├── Your instructions, every time ─────────────── Deterministic
│ ├── One pass ................................ Augmented LLM
│ ├── Ordered steps ........................... Prompt Chaining
│ └── Independent parts ....................... Parallelization
│
├── The AI's judgment, by your method ────────── Guided
│ ├── Sorting inputs into kinds ............... Routing
│ ├── Grading work and sending it back ........ Evaluator-Optimizer
│ ├── Several views to decide ................. Parallelization (voting)
│ ├── Passing to a specialist mid-task ........ Handoff
│ └── Splitting into a known kind of
│ subtasks, with a cap .................... Orchestrator-Workers
│
└── Only the goal can be described ───────────── Autonomous
├── One agent working it out ................ Autonomous Agents
├── Several workers on subtasks it invents .. Orchestrator-Workers
├── Deliberation ............................ Group Chat
└── Peers refining together, with no lead ... Swarm

Then ask a separate question: is there a high-risk action, or one that can’t be undone? If so, add a Human Checkpoint before it. A checkpoint sets how a person takes part (involvement); it never changes the level.

Each pattern maps to an autonomy level by what decides the next step. Where a pattern can sit at two levels, the table says what tips it.

PatternUse whenWhat decides the next step (level)Also called
Augmented LLMOne call, with retrieval, tools or memory, does the jobDeterministicsingle call
Prompt ChainingThe work splits into ordered steps, each checked before the nextDeterministic with coded gates; Guided if the AI judges a gatesequential pipeline
RoutingInputs fall into kinds that each need different handlingGuided if the AI classifies; Deterministic if rules routerouter, dispatcher, triage
ParallelizationIndependent parts can run at once, or several attempts should voteDeterministic when the split is fixed; Guided if the AI judges the voteparallel, fan-out / fan-in, sectioning, voting
Evaluator-OptimizerClear criteria can grade a draft and send it backGuidedreview loop, generator–critic, maker-checker, iterative refinement
Orchestrator-WorkersA lead splits the work, hands out subtasks and combines the resultsGuided when the planning is bounded; Autonomous when it’s open-endedhierarchical, manager, supervisor, coordinator, agents-as-tools
HandoffThe right specialist changes as a conversation or task unfoldsGuidedOpenAI’s “decentralized”
Group ChatThe answer benefits from several agents debatingAutonomousdebate
SwarmPeers refine the work together with no lead, by messaging each other or through shared stateAutonomousdecentralized (the playbook’s earlier name)
Autonomous AgentsYou can describe the goal, not the stepsAutonomousagent loop, gather–act–verify
Human Checkpoint (cross-cutting)A high-risk or irreversible action needs a person’s approvalNo level. It sets involvement (Augmented)human-in-the-loop, approval gate

A person’s review never changes the level, and neither does the number of agents. The workflow’s level is the highest level any of its AI steps reaches.

Whether the workflow runs as a Skill or an Agent is a separate choice. Design makes it by how the workflow runs (whether it decides its path at runtime, uses tools on its own judgment, or runs unattended), never by the pattern.