Engineering Manager. Twelve years in software, the last five leading distributed teams for clients in the US, UK and Europe. Most of my attention now goes to how an engineering team should work when most of its code is written by agents.
One rule: no code is written without a plan file that authorises it, and that applies to me as much as to the agent. Five phases, from an adversarial alignment interview through plan review to execution in a clean context.
It exists because once a team generates most of its code, reviewing the output stops working. A thousand lines of plausible-looking code is exhausting to read, and you approve things you don't fully understand. A two-page spec takes ten minutes, and it's where the decisions actually are. So code review moves upstream and becomes plan review.
I run it on a healthcare platform in Chile that handles the legal import of high-cost medication. The repo shows the method on a synthetic reservation system, since the real domain belongs to a client.
I set the goal and the reasons behind it in as much detail as I can manage, and then I don't police the method. What I review is the result. Engineers who understand why something matters make better calls than the ones I would have made for them, and they make them without me in the room.
I take mandates that arrive as one line with no committed customer yet. At Pariveda that meant a healthcare data squad, a developer bench of up to 20 people and a greenfield MVP, running in parallel. Consolidating the documentation and cutting IAM back to least privilege took monthly bug reports from about 12 to between two and four, over three sprints. The sentiment analysis product went from $80 of Bedrock burned in two days to roughly $2 a day. At Seeed I was effectively the CTO and built the hiring process from nothing: assessment, scorecards, offer negotiation. Four engineers and a PM came through it.
The work I'm proudest of is people getting better. One engineer joined with no technical background at all and was shipping production code and AWS certified inside a year. I coached 12 consultants at Pariveda toward five cloud certifications. Before that I taught about 60 people a night over Discord in a community of 700, and later designed the labs and rubrics as an instructor at BIT Bogotá.
TypeScript, Next.js, Node. Python when the work is data. AWS: Glue, Redshift, Lambda, CDK, Bedrock. Postgres with Prisma, dbt and Airflow. Claude Code, MCP servers, n8n.