Choose a Pattern: Pick by Failure, Not by Name

Read this after the first six chapters.

This page does not explain how a chatbot becomes an agent; Start Here does that. This page only routes failures to patterns.

First Decide Who Controls the Next Step

flowchart TD
  A["Task arrives"] --> B{"Can one model call solve it?"}
  B -->|yes| ONE["Stay with a plain chatbot"]
  B -->|no| C{"Are the steps known upfront?"}
  C -->|yes| WF["Workflow: code controls the path"]
  C -->|no| D{"Does the next step depend on observations?"}
  D -->|yes| LOOP["Agent loop: model proposes, Python executes and limits"]
  D -->|no| PLAN["Planning / Search: design or search the path first"]
  LOOP --> E{"Is one agent overloaded?"}
  E -->|yes| MA["Multi-Agent: split roles or transfer ownership"]
  E -->|no| RISK["Add reliability, retrieval, safety, and eval as needed"]

The map is about control:

Workflow: code fixes the next step.
Agent Loop: the model chooses the next action from state.
Planning / Search: generate or search a path before executing.
Multi-Agent: stop making one agent own every responsibility.

Symptom Table

Failure you see	Start with	Why
A large task drifts	Prompt Chaining	Split one big task into controlled steps.
Different inputs need different paths	Routing	Classify first, then run the right workflow.
Tool-call count is unknown	ReAct	Let observations decide the next action.
Answers sound plausible but are often wrong	Maker-Checker	Generate first, then critique.
You want several candidates to calibrate each other	Voting	Sample multiple answers and aggregate.
Long answers need claim-level fact checks	CoVe	Extract claims, then verify them one by one.
One retrieval pass misses evidence	Retrieval Loop	Read evidence, improve the query, search again.
Answers need citations and an evidence ledger	Agentic RAG	Let the agent decide what to retrieve and whether evidence is enough.
Failure lessons should affect the next attempt	Reflexion	Write task reflections back into memory.
A long report needs section-wise research	STORM	Gather by perspective or section, then synthesize.
The task needs an explicit plan	Plan & Solve	Plan first, then execute.
New information can invalidate the plan	Planner-Executor-Replanner	Make replanning explicit and budgeted.
Tool calls can run in parallel or depend on each other	ReWOO / LLM Compiler	Lay out tool steps before observing or executing.
There are many possible paths to try	LATS	Treat trajectories as a search tree.
The model needs to choose a reasoning strategy	Self-Discovery	Select reasoning modules before solving.
One agent owns too much work	Manager-Worker	A manager assigns owned subtasks.
A specialist agent should be callable as a capability	Agents-as-Tools	Put the specialist behind a tool boundary.
Ownership should change mid-conversation	Handoff	Transfer the task to a better owner.
Several roles need to challenge each other	Group Chat	Put agents in a governed discussion.
A long task stalls or repeats delegation	Magentic Orchestration	Use a task ledger and stall detection.
Tools affect the real world	Policy, Guardrails, HITL	Limit permissions and require approval for risky actions.
You cannot tell whether a change helped	Eval Harness	Run fixed tasks as regression checks.

Cost Rule

Every pattern buys something and charges something.

Pattern family	Buys	Charges
Workflow	predictability and testability	more fixed steps
Agent Loop	adaptation from observations	latency, cost, loop failures
Reliability	more trust	extra model calls and validators
Retrieval & Memory	external knowledge and experience	source quality and context management
Planning & Search	longer task horizon	state, budget, search space
Multi-Agent	role separation and parallelism	coordination and debugging cost
Safety & Evaluation	deployability and regression checks	rules, logs, test sets

If a simpler structure solves the failure, stop there.