Agentic coding, and how it differs from autocomplete.
The difference is the feedback loop. An agent can find out it was wrong.
Agentic coding means a model works toward a goal over multiple turns, deciding what to do next based on what happened last, rather than producing one response to one prompt. The loop is: plan, call a tool, read the result, revise, repeat until done or blocked. The tools are ordinary developer actions such as reading files, searching, running commands and editing. The practical difference from autocomplete is that an agent has a feedback loop, so it can run the tests and react to a failure rather than handing you code it never executed.
Where the difficulty actually lives
It is tempting to think the model is the whole story, but most of what separates a useful agent from a frustrating one is the scaffolding: which tools it has, what it is allowed to do without asking, how results are fed back, how much context is kept and what is dropped. That surrounding machinery is what harness engineering names, and it improves independently of the model.
Vibe coding is a different thing
The two terms get used interchangeably and should not be. Vibe coding describes accepting generated code without closely reading it, which is a stance toward review. Agentic coding describes the mechanism by which the code is produced. You can use an agent and review every diff carefully, and that is neither a contradiction nor unusual.
Why verification is the binding constraint
An agent that can run your test suite can tell whether it succeeded. An agent that cannot is guessing, however good the model. This is why the most valuable thing many teams do before adopting one is make the tests fast and reliable: the agent's ceiling is set by how well it can check its own work.
Autonomy is a dial, not a switch
Serious harnesses offer graded modes: explore and propose but write nothing, propose each edit for approval, accept edits but confirm commands, or run freely. The right setting varies by task and by how well you know the codebase, and treating it as one global preference is how people end up either reviewing trivia or missing something important.
Questions
Related
- What Is the C4 Model?Four levels of zoom on one system. Two of them get maintained; the other two usually do not.
- What Is an Architecture Decision Record?The code shows what was decided. Only the record shows what was rejected, and why.
- What Is Architecture Drift?Nobody decides to let the diagram go stale. It goes stale because nothing in the process notices.
- What Is Diagram as Code?Text in, picture out. Which means the layout engine, not you, decides what it looks like.
Last updated