Insight · July 10, 2026
What is
agentic engineering.
A year ago, AI finished the line you were typing. Now it takes the ticket, plans the work, writes it, tests it, and hands you a diff to review.
01 · The shift
The tool stopped finishing your sentence and started finishing your task.
A year ago, AI in software meant autocomplete. You typed, it suggested the rest of the line, and you stayed in control of every decision. The model was a faster keyboard. You were still the one doing the work.
Now the same tools take a whole task. They plan an approach, edit across many files, run the tests, read the errors, and fix what breaks, iterating toward a working result before they ask you anything. That is not a smarter autocomplete. It is a different unit of work. The unit went from the keystroke to the task, and almost every question worth asking about AI and software follows from that one change.
02 · What it is
Agentic engineering is directing agents, not typing code.
Agentic engineering is the practice of using coding agents as force multipliers under your direction, while you keep full responsibility for architecture, scope, and quality. The agents are goal driven. You give them a target, they plan, execute, test, and iterate with little input between steps, and you act as the supervisor who defines the goal and judges the outcome.
It helps to name the thing it is not. In February 2025 Andrej Karpathy coined vibe coding for the workflow where you describe what you want, accept whatever the model produces, and never really read it. That is fine for a demo or a weekend prototype. Agentic engineering is the version that survives production, because the review never went away. It moved. You scope the task, the agent builds it, and you read the result rather than every keystroke.
The distinction is not about how powerful the model is. It is about where the human stands. Vibe coding keeps you as an ever present conductor. Agentic engineering makes you the person who sets the bar and checks the work against it.
The one line to keep
“Vibe coding accepts the output. Agentic engineering reviews it. The judgment did not disappear. It moved upstream.”
03 · The anatomy of an agent run
Every run is the same shape, no matter the task.
01
Scope
You define done in terms the agent can check. A ticket, a spec, a failing test. A goal it cannot measure produces work you cannot trust.
02
Plan
The agent reads the codebase, decides which files to touch, and lays out the steps before it writes a single line.
03
Execute
It edits across files, runs commands, and moves through the plan without stopping to ask permission at every step.
04
Verify
It runs the tests, reads the errors, and fixes what fails. The loop closes here, inside the agent, not at your keyboard.
05
Review
You read the diff, not every keystroke. Your judgment lands on the finished output, where it counts most.
The beat that separates agentic engineering from everything before it is verify. A prompt produces output and hopes. An agent checks its own work against a bar you set, fixes what fails, and only stops when the bar is met. That is the same discipline behind loop engineering and context engineering. Different names, one muscle.
04 · The delegation gap
Developers use AI everywhere and still finish the job themselves.
Anthropic's 2026 Agentic Coding Trends Report put a number on the strange part. Developers now use AI in roughly 60 percent of their work. Yet they report being able to fully delegate only 0 to 20 percent of tasks. The report calls that space the delegation gap, and it is the central finding, not a sign that adoption is running behind.
Look at what falls on each side. Engineers hand off the work that is easy to verify or low risk. The moment a task turns design heavy or ambiguous, they pull it back to their own hands. The gap between using AI constantly and trusting it to finish the job is not a phase that closes as the models improve. It is the permanent layer where human judgment lives.
That is the real lesson of agentic engineering. It is a supervision discipline. The agent does more of the doing. You do more of the deciding, and the deciding is not going anywhere.
05 · One run that shows the ceiling
Seven hours, no code written by hand.
Kenta Naruse, a machine learning engineer at Rakuten, pointed Claude Code at a large, unfamiliar, multi language open source codebase and asked it to implement an activation vector extraction method. It worked on its own for seven hours in a single run. The result matched the reference method to 99.9 percent numerical accuracy. As he put it, he did not write any code during those seven hours. Across the company, the time to ship a new feature fell from 24 working days to 5.
Read that as the top of the range, not the middle. It worked for two reasons that have nothing to do with luck. The finish line was measurable, numerical accuracy against a known reference. And a person scoped the task and checked the result. Take away either one and you are back to hoping a confident answer happens to be right.
/goal implement the activation vector extraction method so its output matches the reference to within 0.1 percent numerical error. Plan the change across the files it touches before editing. Run the test suite and the accuracy check after each change. Fix what fails. Stop only when the accuracy bar is met.
A scoped goal · the agent runs until the bar is true
06 · What the job becomes
The work moves from writing to scoping and reviewing.
When the agent handles the doing, the role shifts toward agent supervision, system design, and review of the output. The senior skill does not vanish. It moves upstream, into the spec and the diff. You get better at writing a goal an agent cannot misread, and better at reading a change fast enough to keep pace with a worker that produces them faster than you could type.
The bottleneck changes with it. It stops being how quickly you write code and becomes how clearly you can define done, and how fast you can judge whether the work meets it. That is the same shift covered in the ralph loop and burn tokens, not headcount. One operator who is fluent at scoping and reviewing now directs a bench of agents that used to take a team.
None of that lands without taste. An agent will happily ship a plausible answer to the wrong problem. Knowing which problem is worth solving, and whether the output is actually good, is the part that stays human.
07 · When it is the wrong tool
Not every task wants an agent, and not every agent wants autonomy.
If you cannot describe the finish line in terms the agent can measure, it cannot verify its own work, and an agent that cannot verify is just vibe coding with extra steps. Design heavy and ambiguous work still starts with a person. Hand an agent a clear target, not a hunch you have not finished forming.
The cost is real too. A system that plans, runs, and checks itself calls the model many times per task. You spend the run to save the review. That trade pays off on work that repeats or breaks into many verifiable pieces. It is wasteful on a one line change you could make faster by hand than you could brief.
Autonomy is a dial, not a switch. Turn it up on the parts with a measurable bar. Turn it down where a wrong step is expensive or hard to undo, and keep those on a short list you approve yourself. Good agentic engineering is mostly knowing where that line sits.
Closing
The job was never the typing.
Pick one task with a finish line you can measure. Write the goal so an agent cannot misread it. Hand it over, then review only the result. The first time you approve a diff you did not type, the role has already changed.
Anthropic 2026 Agentic Coding Trends Report · Rakuten and TELUS customer stories, Claude by Anthropic · Andrej Karpathy on vibe coding, February 2025 · AGENT 2026, International Workshop on Agentic Engineering, ICSE
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.