Blog
The Reviewer Was Certain the Files Didn't Exist
Two agent runs sharing one working tree don't collide with an error you can see. They produce a correct-sounding report about the wrong state of the world, and by the time you catch it you've debugged the story the race told you instead of the race itself.
The Bar for a Legacy Codebase Is No Worse Than Today
Enforcement a team disables is worse than no enforcement at all, and grandfathering existing debt is what lets a gate survive contact with a real codebase instead of getting disabled by day two.
Eight Tasks Went Out. The Report Came Back With Nine Outcomes
A run that reports success doesn't always mean the work landed. A status line whose counts stop reconciling turns a normal pause indistinguishable from a cascade of real breakage, and that costs more than the bug itself.
The Code Was Never the Hard Part of a Cross-Repo Change
A change whose blast radius is four repos is one unit of intent and four units of bookkeeping. Cross-repo engage and arch headroom close the loop an earlier post left open.
What I Actually Do Now That Claude Writes the Code
After 12 months and 75K lines of AI-generated code, my job looks nothing like it used to. It's harder in different ways.
Your AI Coding Agent Doesn't Know Your Other Repos Exist
Multi-repo is the right architecture. The tooling just hasn't caught up.
Anthropic's Own Teams Prove the Point: Context Engineering Is the Job Now
Anthropic published how ten internal teams use Claude Code. Their number one recommendation is the same thing I've been building around for a year.
The Day I Stopped Thinking of PairCoder as "Claude Code with Guardrails"
For over a year, I described PairCoder as Claude Code with enforcement on top. Then a broken GitHub Action at 2 AM showed me what the project had actually become.
We Used PairCoder to Build PairCoder: 12 Months of Eating Our Own Dog Food
Every major feature in PairCoder has a 'we got burned' origin story. After 400+ tasks and 12 months of dogfooding, here's what survived contact with real work at real scale.
40% of Experienced Claude Code Users Don't Review What the Agent Writes
Anthropic's research confirms what a year of AI code review already taught me: human attention doesn't scale. Enforcement has to be structural.
A 4-Hour Task Takes 10 Minutes: What 400 Tasks Taught Us About AI Estimation
Traditional estimation measures the wrong thing for AI-augmented development. After 400 tasks, we shifted from hours to tokens and discovered the human overhead gap nobody talks about.
Your Skill Library Is a Context Tax
Skill packs and skill managers optimize the wrong thing. A year of benchmarks and one embarrassing drift audit taught us: anything deterministic belongs in code that fails closed, skills are a thin judgment layer capped and enforced, and the test for keeping one is whether deleting it would be noticed.
Why AI Agents Need External Enforcement, Not Better Prompts
Better models don't eliminate the need for enforcement. After 400+ tasks and 71K lines of code, here's why structural constraints beat instruction-following every time.