The Variations
Eight shapes you will meet, from no loop at all to loops made of loops.
Ordered from least to most freedom given to the model. The first four are workflows: code decides the shape and the model fills in steps. The last four are agents: the model decides the shape. Five of the names (prompt chaining, routing, parallelization, orchestrator–workers, evaluator–optimizer) are the ones Anthropic's "Building effective agents" guide uses, which is the closest thing the field has to a canon. One-shot, the loop and plan-and-execute are the industry's usual names. The aliases are what everyone else calls them.
One-shot no loop · 1 pass
Also called: single completion · prompt with context · the packet rule · "just call the API"
Tools run in code first, the model answers once. Your house today. Cheapest, most predictable, and the model can never ask for more.
Prompt chaining fixed steps · N passes
Also called: pipeline · sequential chain · multi-step prompt · LangChain "chain"
Code calls the model several times in a fixed order, each call's output feeding the next, with programmatic checks between. No branching by the model.
Routing one fork · code executes
Also called: classifier + dispatch · intent routing · model router · triage
The model picks a category; code runs the matching path. The model decides which, code decides how. Used to pick a cheaper model or a specialist prompt.
Parallelization fan out · aggregate
Also called: fan-out/fan-in · sectioning · voting · ensemble · map-reduce
Code runs several model calls at once, on parts of a job or on the same job for votes, then merges. Different from parallel tools inside one pass.
The loop ReAct · the ring
Also called: ReAct (reason + act) · tool loop · agentic loop · function-calling loop · "an agent" · autonomous agent
Sections 01 to 05 of this page. The model chooses tools and when to stop; the runtime executes and guards. The base pattern every framework wraps.
Orchestrator–workers a loop whose tools are loops
Also called: multi-agent · supervisor · manager/worker · agent-as-tool · subagents · crew · swarm · "the org chart"
One agent's tools are other agents. The orchestrator decides what to delegate; each worker runs its own loop and returns a result like any tool would. The org poster in Gamble's House, You → GI → the crew, is this shape.
Evaluator–optimizer two loops facing each other
Also called: generator/critic · reflection · self-refine · reviewer loop · LLM-as-judge in a loop
One model call produces, another grades against criteria, and the critique goes back as context until it passes or the breaker trips. Validate, but with a model doing the judging.
Plan-and-execute plan once · loop per step
Also called: planner/executor · task decomposition · to-do agent · hierarchical agent
A planning call writes the step list; a loop runs each step; optionally the planner re-plans when a step fails. Trades freedom for legibility: you can read the plan before it runs.
The line between the two halves is the one that matters for cost and trust. In a workflow, you can draw the shape before it runs, so you know the worst case. In an agent, the model draws the shape as it goes, so the breaker is the only thing standing between you and a surprise bill.