Tools agents can use
MCP servers that connect agents to real systems and useful actions.
Twenty years building production software. At Zapier, I build MCP servers, reusable agent skills, and agents that work in the background. I also lead the security work that defines what those agents can access and do.
the whole assembly line
"Full stack" usually means a frontend and a backend. Mine starts before the code and doesn't stop at the deploy.
Spot the problem worth solving, and kill the ideas that aren't. Decades of tinkering built a good filter for it.
Talk to users, cut to the MVP, write the spec nobody asked for. I've owned discovery through roadmap on cross-org products.
Frontend, backend, data, AI agents, whatever the product needs. Twenty years across Python, TypeScript, Scala, and more.
CI/CD, infra, observability, rollbacks. Deploys should be boring, so I build them that way.
Test it, abuse it, watch real users hit the edges. Testing happens as I build, not in a phase at the end.
Instrument, learn, cut what isn't working, ship the next version. The loop doesn't stop.
the systems around the agents
Useful agents need more than a model. They need tools, clear instructions, and safe access.
MCP servers that connect agents to real systems and useful actions.
Clear instructions and workflows that help agents handle repeat tasks.
Async and cloud agents that do useful work outside a chat window.
Sandboxing, permissions, and secret access. Define what agents can do and where a person must step in.
proof of shipping
Highlights from twenty years of startups, scale-ups, and one very large bird app. Full detail lives in the resume.
Build MCP servers, reusable agent skills, and async and cloud agents on Zapier's internal AI development team. Lead security, privacy, and governance for those agents.
added at Twitter by internal tooling I designed that enabled millions of new push notifications.
data pipeline at Nebula.io after my event-driven ETL rebuild, at 85% lower cost.
ownership of Zapier's entitlements system from discovery to production. Scoped needs across teams and led the build.
on web ads at TapJoy. Ad views rose over 200%.
featured by Apple. 15k copies sold in two weeks, built solo.
built and sold two SaaS companies, InvoiceJournal and InnFirst, from idea to acquisition.
operating principles
I care more about how the product feels to use than how long the feature list is. I sweat the flow, the copy, the empty states.
Developer experience is a product too, and its users are engineers. Fast local dev, boring deploys, tools people actually want to use.
Momentum compounds. Small scopes, tight loops, ship weekly. Shipping is how I find out if I was right.
Twenty-plus years across Zapier, Twitter, and a string of startups, with the numbers attached.
Agents welcome. The resume is machine-readable: GET /resume.json (JSON Resume schema), plus schema.org JSON-LD on every page.
currently shipping
Real products, live on the internet. I designed, built, deployed, and run each one.
A deployment platform built for the agent era. Repos, builds, previews, databases, domains, rollbacks: one control plane that humans and AI agents can both drive safely.
deploymill.comA bank for your ideas. Capture them the second they hit, develop the promising ones, and let the rest sit and compound until they're ready to build.
ZiggyPlus a rotating bench of experiments: MCP servers, AI agents, internal tools, home projects. Something is always mid-build. See the rest on GitHub.
notes from the build
Field notes on building with AI, shipping small, and twenty-odd years of lessons. Published on Medium.
Everyone else builds for the median user. I'm building Ziggy for one person: me.
Meant to be a victory lap: 1,286 commits into a 28-day build. It became a lesson about shipping the thing that actually matters.
Three decades of writing software, and AI still managed to upend how I think about the craft.
A clean, frequent Git history is one of the easiest ways to get better results out of AI coding tools.
More context isn't always better. Relevant, high-quality context is what actually moves the needle with LLMs.
Half a career in Jira and I still write bad tickets. Here's what I'm doing to fix that.
For most of my career I worried too much about scaling. Then I stopped, and started shipping.
Using ChatGPT as an email sounding board, to route around the parts of the interface that got in the way.
More on Medium →