Agentic workflow

An agentic workflow is a process in which AI models and tools run through a defined sequence of steps to reach a goal, such as planning, writing and reviewing a code change. Each step may use a model's judgement, but the path between steps is set in advance.

What it is

Anthropic's Building effective agents draws the line this way: workflows are systems where models and tools are orchestrated through predefined code paths; agents are systems where the model directs its own process and tool use. It names common workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer.

Most practical systems mix both: a fixed outer sequence, with steps that can call tools in a loop.

Why it matters

A fixed sequence is easier to reason about. You know which step runs when, what it receives, and where a person or a check sits. For changes to production code, that predictability is usually worth more than full autonomy.

In AI CodeOps

A session runs one of 16 built-in workflows. The default, Plan, Build, Review & Test, has four steps: the architect plans, the builder writes the change, the reviewer reads it, and the tester checks it. Steps run one after another, and each gets its agent's instructions plus the earlier steps' output.

When a session produces a diff, AI CodeOps opens an approval request. The pull request opens only after a person approves it, and the workspace policy and release gate get their say as well.