An original drawing inspired by Joan Miró, whose work I love. The sun in it sits where your sun is right now; after sunset it turns into a moon.
David Manzanarez
I build distributed systems and agentic workflows for Amazon.com, mostly in Checkout and also across the wider Purchase Experience organization. The systems I build and own handle over a million requests a second and serve hundreds of millions of customers a day.
These days I lead technical work across several Checkout teams and build agents that take on software maintenance. Before that I led the Checkout side of Add to Delivery and a multi-year move from a monolith to Java microservices.
I live in Seattle. Outside software I read about ancient cities and go to stand in what is left of them.
Some things I’ve worked on
Agents for engineering work
2025–present
I designed and built a system that assigns maintenance tasks, such as removing expired feature flags, to agents. Agents are not deterministic: the same task can stall, produce a different change on a second run, or fail halfway through. So the system treats each task as a durable workflow rather than a single prompt. A task is checked against an operating procedure before it starts, then runs as a state machine that records every step. When a run stalls or fails, it picks up from the last recorded step instead of starting over. Outputs are evaluated before a human reviewer sees them, and nothing merges without that review.
I also built Model Context Protocol (MCP) tooling that lets agents investigate checkout failures across purchase-journey data and service logs. Caching and query limits keep those investigations from adding load to production.
Add to Delivery
Add to Delivery lets a customer add an item to an upcoming delivery with one tap instead of placing a new order. I set the Checkout architecture and led execution across more than 20 engineers in four organizations. My part connected delivery data to checkout services and the customer-facing experience, so the button could appear on product, search and reorder pages. I later led its expansion to five European marketplaces.
I was the technical lead for a migration to Java microservices involving more than 30 engineers. I designed core APIs and plugin interfaces and coordinated the rollout. Teams elsewhere at Amazon had to keep shipping features while we replaced the platform they depended on.
Early in the migration I built the framework we used to measure checkout latency. Before the migration, I worked on chaos testing for checkout, protection for one-click purchases, and the framework that formats languages and currencies for every marketplace.
At home
I have a computer at home that runs research agents overnight and emails me a report in the morning. Most of the experimenting is in how to give them useful questions and how to check their answers.
I build my own projects with Claude Code and Codex, often both on the same task. They are a small set of systems for notes, studying, biology and genetics, finance and travel.
The overnight server
The computer is an Ubuntu server I set up with an unattended install. Its disk is encrypted with LUKS and unlocks itself from the TPM at boot, and it can only be reached over Tailscale. At night it runs Claude Code orchestrations with nobody watching, and the morning report is one of them.
Checking an API bill
The first tool built on that server compares the totals on an API invoice with the usage recorded for each request, recomputes what those requests should have cost under prompt-caching prices, and flags the differences. I had a night shift of agents build it, and they left a full audit trail and a write-up of what they learned.
Personal systems
I built eight services for my own use. They include a journal with full-text search, health and biometric tracking, finances, a study system that schedules reviews with FSRS spaced repetition and runs Python in a sandbox, and a hub that signs me in to all of them with Google. They run as Docker containers on DigitalOcean, written in TypeScript with Hono, React and SQLite. I write a spec first and build from it with Claude Code. A template for one service is public on my GitHub.
My genome
I analyzed my own 30x whole genome end to end: variant quality control, ancestry with supervised ADMIXTURE and relatives as controls, polygenic scores, pharmacogenomics, and Y-chromosome and mitochondrial lineage. Along the way I ran into method traps, such as where PCA and admixture analysis disagree and what confounds dating ancestry from tract lengths, and I wrote each one up.
How I see the world
I am an optimist about technology and the trajectory of human civilization. Most people who ever lived spent their lives just above subsistence, facing what Noah Smith calls “the elemental foe”: poverty. That began to change only in the last two centuries. Works like Smil’s Energy and Civilization and Pinker’s Better Angels helped shape that optimism.
None of it was automatic. Things got better because people decided what to build and then built it, and every part of that can still be lost. Technology is not neutral, and it is not dangerous by default. What matters is what we choose to build and whether it actually helps people.
I want to see how far this goes. Think about what it takes to make one modern CPU: hundreds of fabrication steps, an EUV lithography machine that only one company on Earth can build, and a supply chain of mines and factories on several continents. No one person understands the whole thing, and it works anyway. If human history runs long, we are the ancients, and what gets built this century will matter for a very long time.
Interests
I am interested in psychology, biology and astrophysics: why people want what they want, how living things work, what is out there. Genetics, longevity and asteroid mining can keep me busy for a while. Given the time, I would visit laboratories, factories and observatories the way I visit ancient cities. I also travel for paintings, mostly Bosch and Miró.