Case studies from commerce delivery and the agent systems behind it.
Most of it comes from contract delivery work through a digital agency, on direct-to-consumer Shopify builds; two pieces are personal systems I use across all of it. Each is written up as the problem, the calls I made, and what shipped versus what is still unproven. Clients are described rather than named.
On a replatform for a direct-to-consumer pet brand, a team of AI agents handled intake, triage, building and reply drafts. I designed the supervision, the independent review, and the limits on what they could touch.
A system meant to answer clients’ questions about their own stores from three sources at once: our handover documents, the platform’s help articles, and the code of their store. Agents built it; I specified it, verified it, and ended it when it became more machinery than the problem deserved.
One compares a live Shopify storefront with its design file and writes down every place they disagree. The other reads a live site with no usable design file and writes the style guide that should have existed. Neither may change anything it looks at.
A folder of plain Markdown notes that a person and an agent both write into, and a store underneath that turns those notes into something an agent searches at the start of a session. I didn’t write the memory software; I designed the system around it.
I put every connection between my AI assistants and outside systems behind one local gateway, Toolport. I didn’t build it. I chose it, configured it, and decided which of its protections were worth the friction.
For a client retainer, a small app that shows priorities and a rough timeline on swimlanes, kept intentionally disconnected from the ticket systems on either side.
A colleague’s text-only prototype, taken to production: it reads a project’s statement of work and produces the schedule and the tickets, following the agency’s standard structure. It is in daily use.
Client feedback arrived as scattered comments across the messaging tool, the QA platform and email. An agent rebuilds the threads, files itemized tickets, and acts only on work inside the scope it was given.
A chain of agent steps that took product images from background removal to a staging store to production, built around one requirement: every asset reaches the right product.
For twelve years I ran digital projects. Now I mostly design the systems that run them: what the agents do, what they’re not allowed to touch, and how their work gets checked before anyone sees it.
My first real development work, and I was pushed into it: the project hit a difficult stretch, and what started as looking for a way through became me carrying most of the build. The AI worked as a fancy autocomplete, so I didn’t need to know every piece of syntax; the enabling work was still mine.
After managing the design and build of the brand’s US site, I forked it and launched the UK store, moving product data with Matrixify and store configuration through the Shopify Admin API in Claude-driven workflows. The challenge was doing the entire migration in two weeks.
Agents did the intake, the building and the reply drafts; my job became specification and verification. I ran it from an issue tracker instead of a code editor, because the unit of work was a ticket, not a file. It’s the first case study above.
An agent may act where its work can be checked or undone, and has to stop and ask where it can’t.
That rule runs through the agent work here, from the build pipeline’s limits to the feedback agent’s scope. Drafting a client reply is encouraged. Sending it is mine.
The best tool for a job depends on what the job is and on when you’re asking. Right now Claude is the base for reasoning and code execution, cheaper runtimes take the work they suit better, and every assistant reaches outside systems through one gateway.
Claude, through Claude Code and the API, for most reasoning and code; OpenAI and OpenRouter where they fit better. Agents run as a team in an issue tracker, each with a scope and a list of things it may not touch.
Tool connections behind one local gateway, plain Markdown notes as the source of truth, and a memory layer that turns those notes into something an agent searches at the start of a session.
Shopify and Shopify Plus, the apps built around them, and the tools a delivery team runs on.
I’m open to select consulting engagements. For anything the case studies don’t answer, email me.