What is an agent loop?
An agent solves a task by looping: it reads its input, thinks, generates a tool call, and the harness acts — the result becomes the next input, and it loops until done. A tool is a function the model may call. A trace is the recorded log of every step. A guardrail stops an unsafe or runaway action.
INPUT→THINK→GENERATE→ACT↺ result → next input
Agents — they loop until done
the model decides the next move each turn
What is a workflow?
A workflow is a fixed pipeline: the stages and their order are decided by code, not the model. Each stage either runs deterministic logic or calls the model exactly once, then hands off to the next. No looping, no tool choice — the model fills one slot and the harness drives the rest. Predictable and cheap; the trade-off is no autonomy.
CODE→CODE→MODEL→CODEone fixed pass, no loop
Workflows — fixed pipelines
deterministic stages; the model fills one slot