Agentic workflow
A working mode where an AI agent autonomously executes tasks over an extended duration, with access to files, tools, and the development environment.
Definition
An agentic workflow is a mode of collaboration with an LLM where the agent doesn't just generate text — it executes actions: reading and modifying files, running commands, calling APIs, interpreting results, and adapting accordingly in an iteration loop. The fundamental distinction from classic "AI chat" is that the agent has access to tools and can use them autonomously across multiple steps without human intervention at each step.
postcursors perspective
The term is often misused to describe any interaction with an LLM that produces code. A truly agentic workflow implies autonomous execution across multiple steps: the agent reads a failing test, understands why, modifies the code, reruns the test, verifies it passes — without being asked to do each step individually.
In practice
In practice: you brief the agent on a task with a verifiable success criterion ("tests must pass"), and let it iterate. The human intervenes to validate the result and orient the next task — not to supervise each step.
Common misconceptions
- ✗ Calling it 'agentic' when you're just copy-pasting a ChatGPT response into your editor
- ✗ Conflating it with AI autocomplete (GitHub Copilot, Supermaven) — both are useful but radically different in nature