All

Software

Agentic engineering means you review the code, not skip it

Agentic engineering is the practice of directing AI agents through a build while a human oversees and validates the output. Here is how it differs from vibe coding and what it asks of creators who ship their own sites and tools.

AÉNL Editorial••3 min read

Agentic engineering is not a faster way to avoid reading code. According to IBM's explainer on the term, it is the practice of using engineering expertise to orchestrate and oversee AI agents through the software development process. The human stays accountable for what ships. The agent does the typing.

For a musician with a store, a filmmaker with a portfolio site or a YouTuber running a small tool stack, that distinction decides whether an AI-built project holds up in month three or quietly breaks during a launch.

What the term actually separates

The label was coined by Andrej Karpathy, the OpenAI cofounder, and IBM's summary of his framing is that agentic programming is a tool, not the force that builds the entire codebase end to end. That is a deliberate contrast with vibe coding, where you describe what you want, accept what comes back and move on.

IBM draws the contrast on one axis in particular. Vibe coding is aimed at speed and experimentation. Agentic engineering is aimed at production-quality engineering. Both use the same underlying models. The difference is the process wrapped around them.

IBM also breaks the two words apart. "Agentic" means an orchestration of agents writes the code while the human developer oversees and validates the output. "Engineering" means there is a level of expertise required to use those workflows for meaningful code production without jeopardizing quality. Remove either half and you are back to something else: unattended automation, or ordinary hand coding.

Why this matters if you are not a developer

Most creators are not employed as engineers, yet many now maintain real software: a WordPress site with custom plugins, a webhook that routes a Stripe or Square sale to an email list, a script that renames and sorts footage, a bot that posts release links. These are small systems, but they touch money and audience data. A bug in one is a business problem, not a learning exercise.

The agentic engineering framing is useful here because it tells you where your effort should go. If the agent writes the code, your job moves to specification, review and verification. That is a skill you can build without being able to write the code from scratch, but you cannot skip it.

The mechanism: why oversight is the product

An agent works in a loop. It reads a goal, plans steps, edits files, runs commands, reads the result and tries again. Each pass can compound a wrong assumption made in the first one. A model that misreads your brief at step one will produce clean, confident, well-formatted code that solves the wrong problem, and it will keep building on it.

That is why the human role in IBM's definition is oversight and validation rather than prompting. Prompting is an input. Validation is a check against reality: does the checkout complete, does the form send, does the page still load on a phone, does the automation fire once and only once.

A practical version of this for a small operator looks like four habits.

Write the brief like a spec

State the outcome, the constraints and what must not change. "Add a booking form to the contact page without touching the header or the existing analytics tags" gives an agent boundaries. "Make my site better" gives it permission to rewrite things you did not ask about.

Keep changes small and reversible

Work in version control, and ask the agent for one change at a time. A single reviewable diff is something a non-specialist can still read for surprises, such as a deleted file, a new dependency or a hardcoded key. A sprawling rewrite is not reviewable by anyone.

Define the test before the build

Decide how you will know it works before the agent starts. For a lead form, that is a test submission arriving in your CRM with the right fields. For a video pipeline script, that is a known folder of clips producing the expected output names. If you cannot describe the check, you cannot validate the result.

Treat secrets and money paths as human territory

API keys, payment logic and anything that emails your audience deserve a manual read every time. These are the places where a plausible-looking mistake costs the most and is noticed the latest.

Where it fits in a creator business

The honest use case is narrow and valuable: repetitive, well-defined work that you would otherwise pay for by the hour or never get around to. Site maintenance, small plugins, content-repurposing scripts, link-in-bio pages, and glue between the tools you already pay for are the typical candidates. Original product strategy, brand positioning and creative direction are not, and an agent that is good at code is not thereby good at those.

If the project is a public-facing site, the foundation still matters more than the tooling. A stable host and a clean WordPress install give an agent less to break, and a provider such as Bluehost is one option for that layer. The agent can edit what you give it, but it cannot make an unreliable host reliable.

If the project is an automation that touches leads or sales, the question to settle first is scope: what should be automated, what should stay manual, and what happens when a step fails. That is a design conversation before it is a coding one, and a short call such as a 30-minute consultation with AÉNL Agency is a reasonable place to pressure-test the plan before an agent starts building it.

The counterargument, and where it is right

The fair objection is that this framing slows people down. If a model can generate a working page in two minutes, a review step feels like friction. For a throwaway landing page or a one-off script you will delete on Friday, that objection wins, and vibe coding is a fine way to work. IBM's own comparison implies as much by placing speed and experimentation on the vibe coding side.

The line to draw is persistence and consequence. Code that will run unattended, handle other people's data or sit in front of a launch deserves the engineering half of the term. Code you will look at once and discard does not. Most problems come from using the first approach on the second kind of project, because the code looked finished and nobody checked.

What to do this week

Pick one small system you already run, such as a form, a redirect, a posting script or a file-sorting job, and sort it into one of two bins: disposable or durable. For anything durable, put it in version control, write down the one-sentence test that proves it works, and make your next agent-assisted change against that test.

That is the whole practice at small scale. The agent does the typing, you own the spec and the check, and the project stays something you can explain to the next person who touches it, including yourself in six months.

For more articles on AI and the business side of creating, the AÉNL Agency newsletter is the place to follow along.

Some links in this article are affiliate links. If you buy through them we may earn a commission, at no extra cost to you.

Bluehost
Square Point of Sale

Make your work stand out with systems built by indie engineers and operators.