What is agentic engineering?
Agentic engineering is building software by directing AI coding agents and taking responsibility for what they produce. You set the goal, supply context and constraints, watch the agents work, review the result and keep the checks that catch mistakes. Andrej Karpathy proposed the name in February 2026 to separate professional work from vibe coding.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
You decide what to build and write it down: the goal, the constraints, the rules of the codebase. Files like AGENTS.md or CLAUDE.md hold the standing ones.
You hand the task to one or more coding agents with the tools they need: the repository, a shell, the tests, a browser.
The agents plan, edit and run things in a loop. You watch for wrong turns and step in before they get expensive.
You review the result the way you would review a colleague's pull request: read the diff, run it, ask why.
You keep the checks that do not depend on anyone's attention: types, tests, linters, review gates. They let you trust code you did not type.
Karpathy's framing: you are not writing the code directly 99% of the time; you are orchestrating agents and acting as oversight.
Source: Karpathy's agentic engineering post, as quoted by Glide
Agentic engineering versus vibe coding
Karpathy coined vibe coding in February 2025 as a joke about his own weekend habit: accept everything, read nothing. A year later the term had swallowed every kind of AI-assisted work, including the serious kind, so in February 2026 he offered a replacement. "Agentic" because the default has flipped: you are not writing the code directly most of the time, you are orchestrating agents that do and acting as oversight. "Engineering" because, in his words, there is an art and science and expertise to it, something you can learn and get better at.
The two terms describe the same tools at the same moment in history. The difference is whether a person stands behind the result. In vibe coding nobody does, on purpose. In agentic engineering someone does, and arranges the work so that standing behind it is possible: small tasks, written rules, tests the agent has to pass, diffs a human can read.
Example: a feature shipped by agents
A team needs a new export format. The engineer writes a short spec: what the file contains, what the edge cases are, which existing module it should mirror. The repository already has a rules file telling agents how tests are run and what not to touch. Two agents work in parallel in separate worktrees, one on the exporter and one on the tests. When they finish, the type check and test suite run without anyone asking. The engineer reads the two diffs, pushes back on a helper that duplicates one already in the codebase, and merges.
Count what the engineer typed: a spec, a rules file months ago, and one review comment. Count what they decided: everything. That ratio is the job.
Why it matters
Models got good enough that typing is no longer the bottleneck. Judgement is. The engineer's time moves upstream into deciding what to build and how to constrain it, and downstream into verifying what came back. What disappears is the middle, where most of a programmer's hours used to go. Teams that notice this reorganise around it: better specs, more automated checks, smaller changes, and reviewers who read for intent rather than syntax. Teams that do not notice get vibe coding with a job title.
Common mistakes
- Reviewing less because the agent is confident. Agents report success in the same tone whether or not they verified anything. Trust the tests, not the summary.
- Skipping the written rules. Every agent run starts from zero. If the conventions live only in people's heads, each run rediscovers or ignores them. Put them in the rules file the agents read.
- Delegating the decision along with the typing. An agent can propose an architecture. Choosing one is the part that needs a person who will still be here next year.
- Measuring output by lines. More code is now free. The scarce thing is code that was needed.
FAQ
- How is agentic engineering different from vibe coding?
- Same agents, opposite stance on review. Vibe coding accepts the output and judges it by whether it seems to run. Agentic engineering treats the output as a draft from a fast but fallible colleague: it gets read, tested and questioned, and the engineer is accountable for it. Karpathy coined both terms, a year apart.
- Do you need to be a programmer to do agentic engineering?
- Yes, in the sense that matters. You need to judge whether a change is correct, spot when an agent is heading somewhere bad, and know what a good test looks like. You type far less code, but the knowledge that lets you read code is what the work rests on.
- What skills does an agentic engineer need?
- Writing clear briefs and codebase rules, choosing what to delegate and what to keep, reading diffs quickly, designing checks the agent cannot talk its way past, and knowing when to stop a run. Karpathy's point was that these are learnable skills with their own depth, not a shortcut around skill.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.