Agentic Workflow

An agentic workflow is the unit of work in the Agentic SDLC. Where a manual prompt automates a single keystroke-level task for one developer, a workflow automates a whole task across the lifecycle: read the failing build, find the cause, write the fix, open the change, and wait for sign-off. It is the difference between asking an agent to do something and having a defined process that does it every time the triggering event occurs.

The shape matters. There is a real distinction between a static workflow, where tools are wired through fixed code paths, and an agent, which directs its own process at runtime. An agentic workflow deliberately combines both. The sequence of steps is defined and repeatable, which is what makes it predictable and governable, while each step is an agent that reasons and uses tools to decide how to get its result. Structure lives at the workflow level; autonomy lives inside the steps. Many workflows running together, each owning one concern and reacting to its own triggers, are what orchestration coordinates across the organization.

A pull request opens and its CI run fails on a broken test. Instead of an engineer noticing the red check, reading the logs, and debugging by hand, the failure is an event that triggers a Fix CI workflow. The first step clones the branch. The next is an agent that reads the CI run and job logs, identifies the root cause, and produces a structured fix plan, without touching any code. That plan flows into the next step, where an agent applies the changes, verifies them, commits with a clear message, and pushes to the branch, then comments on the PR with what it fixed. The engineer reviews the result and approves the merge. The same workflow runs on the next failing build, and the next, without anyone re-explaining the task.