What Agentic Development Can’t Do for You

Once a technical lead gets comfortable with how agentic development works, the next question is almost always the same: what can't it do?

The answer we give tends to surprise people. With a solid plan and an experienced technical lead behind it, nearly any development task can run through AI agents. Theme work, integrations, checkout extensions, data migrations, analytics instrumentation. If the work can be clearly defined, agents can produce it.

What agents can't do is the work that happens before a task is defined and after the code is written. That's the judgment layer, and in our experience it's where engagements succeed or fail.

Production and Judgment Are Different Jobs

It helps to separate two kinds of work that usually get lumped together as "development."

Production is turning a well-defined requirement into working, tested code. Judgment is everything around it: deciding what the requirement should be, how the pieces fit together, what done looks like, and whether the result is actually right.

For most of the history of software, the same people did both, and production consumed most of their time. Agentic development changes that ratio. Production gets dramatically faster. Judgment doesn't, and it now carries more of the weight. Our agentic development model is built around that split: agents handle production, and senior engineers own the judgment.

Here is what that judgment layer includes.

1. Deciding What to Build

Agents implement what they're asked to implement. They don't know your margin targets, which customer segment matters most this quarter, or which promotion your finance team's reporting depends on. Deciding what a feature should do, and why, is a business and experience question before it's a development question.

A vague ticket produces confident, plausible output that solves the wrong problem. That's true of human developers too. The difference is that an agent gets there faster, so the cost of an unclear requirement shows up sooner.

2. Making Architecture Decisions

Where data should live, how systems talk to each other, whether a requirement belongs in a Shopify Function, an app, or the theme. These choices have consequences that outlast any individual ticket, and they depend on trade-offs that aren't written down anywhere: the future roadmap, the team's skills, the platform's direction, and cost.

An agent can implement any of those options well once the decision is made. Making the decision is a senior engineering call, and it should stay one.

3. Defining What "Done" Means

The first step in our workflow is scoping each task into defined inputs, acceptance criteria, and measurable outcomes. That step is entirely human, and it determines whether everything after it works.

Compare two tickets. "Make checkout faster" is a goal. "Bring mobile LCP on the cart and checkout templates under 2.5 seconds without removing the upsell module" is a task. An agent can do excellent work on the second. On the first, it has to guess what you meant.

When a task resists that kind of definition, it's a signal the work needs more human thinking first. That isn't a limitation of agents specifically. A ticket nobody can define isn't ready for a human developer either.

4. Knowing When the Output Is Wrong

Our agents succeed on 82% of the tasks they're assigned. The other 18% is the reason a person who understands the work reviews every change.

Generated code is rarely wrong in obvious ways, because obvious problems fail automated checks. The failures worth worrying about are subtle: code that passes the tests but solves an adjacent problem, handles an edge case by avoiding it, or introduces a pattern that conflicts with the rest of the codebase. Catching those takes someone who knows the codebase and the business well enough to notice. We walked through the full set of review and QA gates in our post on whether AI-written code is safe to ship.

5. Supplying the Context Nobody Wrote Down

Every mature ecommerce codebase has history. A workaround that exists because of an integration quirk from four years ago. A field that looks unused but feeds a report the finance team runs every month. A template that's structured strangely because a merchandiser depends on it working that way.

Agents know what's in the code, the documentation, and the ticket. They don't know what's in people's heads. Part of a technical lead's job in this model is getting that knowledge into the plan before an agent starts, which is also why discovery matters as much as it ever did.

What This Means for Your Backlog

The practical version of all this is a readiness check rather than a list of off-limits work. A task is ready for agentic execution when it has:

• A clear requirement that someone has decided on
• Acceptance criteria specific enough to test against
• Known constraints, including the ones nobody documented
• A reviewer who understands the codebase and the business

Most backlogs have plenty of work that already meets that bar. The rest usually needs a conversation, not a different delivery model. In our engagements, that triage is part of the work, and front-loading the judgment is what makes more of the backlog eligible.

The Red Flag to Watch For

If an agency pitches agentic development as a way to replace engineers, be careful. The model works the other way around. Agents absorb production, which makes judgment more central to the engagement. A staffing plan that entirely removes the engineering layer to make room for agents has the economics backwards.

This can be a recipe for disaster, as engineers are still necessary to provide the proper strategic direction and QA needed to deliver impactful projects that don't fall apart two weeks from when they go live.

The Short Version

Agentic development can take on almost any development task, as long as the judgment around that task is done well. Deciding what to build, choosing the architecture, defining done, reviewing the output, and supplying the context nobody wrote down all stay with experienced people.

As production gets cheaper, that judgment is most of what you're really paying for. If you want a read on how much of your backlog is ready for agentic delivery today, and what it would take to get the rest there, we're glad to walk through it.