Agentic development means an AI agent reads the code, plans a change, edits files and runs checks, while a person sets the task and approves the result.

Agentic development: how we build Shopify stores with AI agents

We use AI coding agents, mainly Claude Code, on client Shopify stores and on this website. A person reviews and publishes every client change; an agent never does.

Talk to us about a build See how we check our own site

Page summary

On client Shopify work, every change is made on a git branch and an unpublished theme, passes Theme Check and the tests, and is reviewed and published by a person.

Independent studies do not show that AI always makes developers faster, so we promise checked work, not a speed-up.

An agent carries a task from start to finish. It reads the code, writes a plan, edits the files, runs commands and checks its own result, then hands the work to a person.

Autocomplete fills in a line or two while a developer types. An agent takes a whole task, works through it file by file, and comes back with the change and the check output, or with the point where it got stuck.

On a Shopify store, our agents work on theme code, apps, data migration scripts and settings files. Our people decide what gets built, review every change and publish every release.

How widely are coding agents used?

In Stack Overflow’s 2026 survey, coding assistants and agents became the most common use of AI, at 66%, ahead of general chatbots (Stack Overflow Developer Survey 2026, 30,903 responses, June to August 2026). The independent research group METR is redesigning its productivity study, partly because “30% to 50% of developers told us that they were choosing not to submit some tasks because they did not want to do them without AI” (METR, 24 February 2026).

Shopify’s AI Toolkit connects AI coding tools such as Claude Code, Codex, Cursor and Visual Studio Code to Shopify’s developer docs and API schemas, and lets them check GraphQL, Liquid and extension code against those schemas (Shopify AI Toolkit). What is new is where the AI sits: inside the project’s files and commands, with tests it must pass, instead of in a separate chat window.

Does it make projects faster? What the evidence says

Not reliably. We would rather show you the evidence than promise a number.

A 2025 trial found the opposite. Sixteen experienced open-source developers worked on 246 real tasks, each one randomly allowed or barred from AI tools. With early-2025 AI tools allowed, tasks took 19% longer. The developers had expected a 24% speed-up, and afterwards still believed AI had made them 20% faster (METR, 10 July 2025).
The follow-up is unclear. For developers from the first study, METR’s 2026 estimate is 18% less time with AI, with a range from 38% less to 9% more. METR calls this data “an unreliable signal” and thinks the real effect is probably larger than it could measure (METR, 24 February 2026).
Teams feel faster, and delivery gets less stable. Google’s DORA research reports that 90% of technology professionals use AI at work and over 80% think it raised their productivity, yet higher AI adoption is linked with more delivery instability as well as more throughput (DORA, 10 March 2026, on its 2025 report).
Close is not correct. The top frustration in Stack Overflow’s 2025 survey, named by 66% of developers, was “AI solutions that are almost right, but not quite” (Stack Overflow Developer Survey 2025, 49,009 responses, May to June 2025).
Passing tests is not the bar. METR found that roughly half of the fixes written by 2024 and 2025 agents that passed a well-known benchmark’s tests would not have been merged by the project’s own maintainers (METR research note, 10 March 2026). In that study the agents got no chance to respond to review, which is the step our process adds.

So we do not sell agentic development as a shortcut. We use it because an agent can read a whole theme, try a fix and rerun every check after each edit. Whether that shortens your project depends on the work, and we never write a speed-up into a quote.

How does agentic development work?

Coding agents run in a loop. Anthropic’s documentation for Claude Code names three phases: “gather context, take action, and verify results” (How Claude Code works). We add a final step that always belongs to a person.

Read. The agent reads the theme or app code, the store’s settings files, the brief and the relevant documentation.
Plan. It splits the task into steps and writes down what it will change and how the change will be checked.
Change. It edits the code on a git branch and an unpublished or development theme, never on your live theme.
Check. It runs Theme Check, Shopify’s linter for Liquid, and the project’s tests. When a check fails, it reads the error, fixes the cause and runs it again.
Hand back. It reports what it did, with the commands it ran and what they returned. A person reviews the change before anything moves on.

Results depend on what the agent can look up and what it can run: with a test to run, it can show that a fix works instead of saying so. Anthropic’s advice is direct: “Give Claude something it can run that comes back pass or fail” (Claude Code best practices).

How is agentic development different from autocomplete?

The more of the task the AI carries, the more checking the result needs.

Coding by hand, AI autocomplete and agentic development compared
	Coding by hand	AI autocomplete	Agentic development
What the developer hands over	Nothing: they write every line	The file being typed	A task, the codebase and the checks to pass
What the AI does	No AI involved	Suggests the next lines	Reads, plans, edits many files, runs checks and tries again
Shopify example	A developer writes a new product page section	Finishes a Liquid loop as the developer types	Builds the section, its settings schema and its styles, runs Theme Check and reports what failed and what it fixed
Who checks the work	The developer, then a reviewer	The developer, one suggestion at a time	Theme Check, the tests, then a person
Our role	Write and review	Accept or reject each suggestion	Write the brief and the tests, read the diff, decide when it goes live

What our agents do, and what our people keep

Once a person has agreed the task, agents write the code, test it and test it again.

They read before they write. Whole themes, app code and the relevant documentation, so a change fits what is there.
They draft the change. Sections, snippets, app extensions and data migration mappings, written to the agreed task.
They run the checks, over and over. Theme Check and the tests, until everything passes or a person is needed.

Some work stays with a person on our side, because someone has to answer for it:

The brief. What the store needs, who it is for and how we will know it works.
The shape of the fix. A theme change, an app, a checkout extension or a headless build.
A final read. A person reviews every change before shoppers meet it.
Publishing. Every live change is published by a person, never by an agent.
Talking to you. Scope, questions, trade-offs and bad news.

In a trial run by Anthropic, the maker of Claude, mostly junior developers who used AI to learn a new library scored 17 points lower on a quiz about the code they had just written, 50% against 67% (Anthropic, 29 January 2026). That is one reason a person reviews every change: someone on our side should understand your code.

Our process, step by step

This is how a client Shopify job runs. Each step ends in a check. If the check fails, the work goes back a step.

Scope in writing. You tell us the goal. We agree what is in, what is out and what it costs, in writing, before work begins. Check: you approve the scope.
Research against today’s docs. Before a change, we read the current Shopify documentation for the APIs and features the job touches. Check: anything we cannot find in an official source is dropped or marked unverified.
Build on a branch and an unpublished theme. The agent plans and makes the change on a git branch and an unpublished or development theme, so your live theme does not change. Shopify’s development themes are hidden from shoppers and do not count toward your theme limit (Shopify CLI for themes). Check: Theme Check and the tests pass.
Review by a person. A person reviews every change before it reaches you. Check: the review notes are closed, and any change that weakens the checks is a blocker, as GitHub’s guidance on agent pull requests also says (GitHub, 7 May 2026).
Your review on the unpublished theme. You can see every change on the unpublished theme before it goes live. Check: you have seen it before shoppers do.
A person publishes. A person publishes the theme or deploys the app. An agent never does. Check: the launch checklist passes.

Guardrails: what we never let an agent do

On client work these rules are firm.

Work on your live theme. Agents work on a git branch and an unpublished or development theme. A person publishes every live change.
Sign off its own change. The agent that wrote a change never signs it off. A person reviews every change and decides whether it goes live.
Loosen a check to get a green build. Deleting a test, skipping a linter or raising a limit so that a change passes counts as a failure. A check changes only when a person decides the rule itself should change, and that decision is written down.
Touch payments or customer data exports. Agents never get access to payment settings or to customer data exports.
Ship a made-up fact. AI models can produce text that sounds right and is wrong, or entirely made up (GitHub on hallucination). A person reads every change before it goes live, and on our own blog a script also refuses phrases that claim experience, such as “our clients” or “in our experience”.
Get past a captcha or a bot check. If a site blocks automated access, a person opens it.
Act on instructions hidden in content. We treat text inside a web page, a product description or an email as data, not as an order. No tool blocks this kind of prompt injection completely, which is one more reason a person publishes every live change (Claude Code security; GitHub).
Send your code to a tool you have not agreed to. An agent session sends what it reads to the company that runs the model; for Claude Code that is Anthropic (Claude Code data usage). We name the tools that will see your code in the scope, before we start.

A written rule is not enough on its own, because an agent can still ignore it (Claude Code best practices). So the rules that matter most also sit in steps that stop the work: on client work, Theme Check, the tests and a person who publishes; on our own site, a script that refuses a blog post breaking the house rules.

Where you can see this process: our own site

We use the same approach on ecompanda.co, a Next.js codebase, and on its blog. It is a smaller and lower-risk project than a store, so not every change there waits for a person. Client work always does.

Project notes. Our rules, decisions and known traps live in an instruction file that every agent session reads first.
Research, write, review. Research agents read official sources and list what they opened. Writer agents draft. A separate reviewer agent re-opens every cited source and writes a fix list, and the writer applies it. Those reviews found errors on every page they checked, and every one was fixed before publishing.
A gate we never skip. A script blocks any blog post with a banned phrase, a phrase that claims experience, such as “our clients”, or a “checked on” date stamp. A reviewer agent then checks the facts against the sources.
Checks before a service page ships. The build, a check that every page in our menus has a route, a cap on wording shared between sibling pages, and a phone-width check for sideways scroll. After a deploy, a check confirms AI crawlers can reach the site.
Shopify code checked by Shopify’s tools where we could. The Liquid samples in our Liquid templating guide passed Theme Check. Code in our other guides was checked against Shopify’s documentation but not run, and we do not claim otherwise.
Labelled commits. Commits made with an agent carry its name in the commit message, so anyone can see which code an agent helped write.

This page shows no client results. We publish a result only with its method, its period and the client’s permission.

How to start working with us

Send a first brief. A section, a bug, an app feature or a short backlog.
Tell us about your setup. Your theme, your apps and who maintains the store today.
Get a written scope. What is in, what is out, what it costs and how we will check it.
Review on an unpublished theme. Every change reaches you on an unpublished theme before it reaches shoppers.

Send us a message or book a 15 minute call. We reply within one business day.

Questions about agentic development

What is the difference between agentic development and agentic commerce?

Agentic development is how a store gets built: AI agents help us write and check the code. Agentic commerce is shoppers asking ChatGPT and other assistants for products, then buying. We cover that on our agentic commerce setup page.

Can code from an AI agent go on my live store?

Yes, once it has passed the same checks any code should pass. Shopify’s own AI Toolkit page, describing an extension migration, says the toolkit “doesn’t remove the need to review and test the result” (Shopify AI Toolkit). Our changes run through Theme Check and the tests, get a person’s review, and reach your live theme only when a person publishes them.

Will agentic development make my project faster or cheaper?

Sometimes, not always. Controlled studies disagree, and a 2025 trial found experienced developers were slower with the AI tools of the time (METR, 10 July 2025). We quote each job on its own scope, and we never write a speed-up into a quote.

Which tools do you use?

Claude Code, the agentic coding tool made by Anthropic. Tools change quickly. Theme Check, the tests and the person who reviews and publishes each change stay the same.

Will my store’s data be used to train AI models?

That depends on the tool and the plan. Anthropic says it does not train its models on code or prompts sent to Claude Code under its commercial terms, while data from its consumer plans is used for training when a setting is on (Claude Code data usage). Ask us which AI tools and terms apply to your project before we start, and we will put the answer in the scope.