Loop Engineering vs Graph Engineering: Designing AI Agent Control Flow

Loop engineering is the design of the repeated cycle a single agent runs: gather context, plan, act with tools, verify, and repeat until a stop condition is met. Its core idea is to loop on evidence, such as passing tests or a validated citation, rather than on the agent's own claim that it is done. The verifier and the stop rules are usually the hard part.

Graph engineering makes the workflow structure explicit. Steps are nodes, transitions are edges, and shared state flows between them, with branching, parallel fan-out and fan-in, retries, human approval points and durable checkpoints defined up front. Each node can itself be a loop, so graphs wrap loops rather than replace them.

A loop is flexible and quick to build: the model decides each next step, which suits open-ended tasks with few tools, but control, cost and debugging get harder as scope grows. A graph trades upfront design for bounded paths, per-step prompts and models, parallelism, auditability, durable state and approval gates, at the cost of more prompts, a state schema and new failure modes. Most tasks are one well-scoped loop; move to a graph when work splits into specialties, needs parallel-then-join, or must be auditable.

Durable execution, observability and open protocols for production agents.

Reasoning and multi-agent research.

Neighbouring pages on this site.