What is agentic coding?
By Sandeep PandaCo-founder, Bug0 · Building FactoryKit
The axis that matters in AI-assisted development is who holds the loop. With autocomplete, you do; the model suggests. With agentic coding, the agent holds the loop (plan, edit, run, read the error, fix) and you hold the outcome.
Definition
Agentic coding is a way of building software where an AI agent carries out the implementation: it plans the change, edits files, runs commands and tests, and iterates until the task is done, while a human defines the task and reviews the result.
The autonomy ladder
- Assistants. Autocomplete and chat. The human types, the model accelerates. Attention required: constant.
- In-IDE agents. The agent edits and runs code in your editor while you watch. Attention required: supervisory.
- Background coding agents. The agent works in an isolated environment and returns a finished pull request. Attention required: at review.
Each rung changes the unit of work: from keystrokes, to sessions, to tasks. The economics change at the top rung, because a person can review many tasks in the time it takes to supervise one session.
Agentic coding vs vibe coding
Andrej Karpathy coined “vibe coding” in early 2025 to describe building by accepting whatever the model produces without reading the code. It is a real and useful mode for prototypes and throwaway projects, and it is defined precisely by not inspecting the output.
Agentic coding is the opposite contract. You still do not type the implementation, but ownership is retained: the task has acceptance criteria, the change runs the repo’s checks, and a human reviews the pull request. Vibe coding optimizes for not reading the code. Agentic coding optimizes for not writing it, while keeping every guarantee that reading used to provide.
What changes in a team’s workflow
Task definition becomes the scarce skill. A well-specified task (what to change, and how to prove it works) determines output quality more than the agent brand does. Anthropic’s own guidance on working with Claude Code makes the same point from the tool side.
Review becomes the job. When implementation is delegated, engineering attention concentrates on two stations: the spec going in and the evidence coming out. Teams that skip this re-read every diff at human speed and conclude, correctly for their setup, that the agent saved nothing.
Where agentic coding breaks
- Underspecified tasks. An agent resolves ambiguity somehow; without acceptance criteria, “somehow” is a coin flip.
- No runnable checks. Agents iterate against feedback. A repo with no lint, no build, and no tests gives the loop nothing to push against.
- Unsafe environments. An agent holding production credentials is an incident report with a delay timer. Isolation and scoped credentials are prerequisites (coding agent security).
None of these are model problems. They are factory problems, which is why we built the line around the agent instead of betting everything on the agent itself (the AI software factory).
FAQs
What is agentic coding in simple terms?
You describe a change; an AI agent plans it, edits the code, runs the tests, and iterates until it is done; you review the result. The agent does the implementation, you keep the ownership.
Is agentic coding the same as vibe coding?
No. Vibe coding means accepting AI output without reading it, which is fine for prototypes. Agentic coding keeps verification and human review, which is how teams use agents on production code.
Is GitHub Copilot agentic coding?
Copilot’s autocomplete is assistance, not agency. Copilot’s coding agent, which takes an issue and opens a pull request, is agentic coding in its background form.
Do you need tests for agentic coding?
You need runnable checks of some kind. Agents improve sharply when they can execute lint, builds, or tests and iterate on failures. FactoryKit runs each repo’s own checks with up to 3 fix attempts.
What is the difference between agentic coding and a background coding agent?
Agentic coding is the practice. A background coding agent is the deployment shape where the agent works away from your editor and returns a pull request.
About the author
Co-founder, Bug0 · Building FactoryKit
At Bug0, Sandeep architects Passmark, the open-source AI regression testing framework. He previously co-founded Hashnode and scaled it to millions of developers as its CTO. He wrote AngularJS: Novice to Ninja, co-wrote Jump Start HTML5, and thinks mostly about agentic testing and the infrastructure that makes AI-written code safe to ship.
See a software factory run on your repos
A demo is a working session: your repos, a real task from your backlog, a finished pull request.
Book a demo