blog
essays on ai engineering, agents, rag pipelines, distributed systems, and what actually breaks when you ship production software.
context is the new compute
Apr 2026in march 2026 anthropic shipped a 1 million token window at standard pricing and inference costs collapsed 240x in 18 months. compute is cheap. context is expensive. here's the discipline that actually separates production agents from demos in 2026.
/7 min read
the 25% agent: designing for the reliability we actually have
Apr 2026a public benchmark ran a 6-task crm agent flow 10 times and hit 25% end-to-end success. meanwhile 57% of orgs claim agents in production. that gap is the most important engineering problem in 2026. here's how i design for the reliability we actually measure.
/9 min read
the code that outlasts you isn't the code you wrote
Apr 2026stewardship is the engineering virtue nobody budgets for. but the thing that actually outlasts you isn't the code, it's the constraints, the names, the data model, and the 'why' doc. and the best act of stewardship is usually deletion.
/8 min read
what i delegate to ai (and what i don't)
Apr 2026ai writes code 10x faster than you do, but you read code at the same speed. here's how i actually work with agents on production systems: what i delegate, what i keep, the verification scaffolding that makes ai useful, and the dumb mistakes that cost me real time.
/7 min read
shipping is the easy part: building software you can actually operate
Apr 2026every engineer celebrates the launch and dies in the maintenance. here's the playbook i use to ship features i can actually live with: the runbook test, the maintenance multiplier, the second-day features, and how to write code for the version of you who has to debug it at 2am.
/8 min read
the wild 20%: where engineering becomes a bet on taste
Apr 2026the boring 80% is about discipline. the 20% is about taste. here's how i decide which non-obvious features to ship, when to bet against best practices, and what the wild calls in regent taught me about earning the right to take risks.
/7 min read
the boring 80%: a tactical playbook for engineering foundations
Apr 2026most engineering 'foundation' advice is fluff. here's the tactical playbook i use to build production ai systems without over-engineering: what to instrument first, the contracts that prevent 90% of bugs, the 30-minute rule, and how to know when foundations are a trap.
/7 min read
building regent, an ai executive assistant from scratch
Apr 2026how i built a production-grade saas that replaces human executive assistants with 3-stage ai email pipelines, RAG memory injection, multi-model routing, and 4-tier stripe billing.
/3 min read