Articles
Notes from shipping software with AI
For engineering leaders who want AI-assisted delivery they can check. Written from our own work, not from the sidelines.
Start here
Review capacity
The review queue is the bottleneck AI coding created
Why more code did not mean more shipped, and where the queue actually forms.
Instruction drift
Agent instruction drift in enterprise AI teams
How every developer ends up with a different AI, and what a shared rule set changes.
What was tested
AI coding checkpoint contract: make checks reproducible
What “tested” should mean before a pull request is allowed to say it.
Latest
-
Does AI coding make developers faster? The current evidence is mixed
AI coding productivity is conditional, not one number. Field experiments report more completed tasks, a bounded Copilot experiment reported faster completion, and…
-
The lifecycle cost model for almost-right AI code
Almost right AI code creates a lifecycle cost that starts after generation. Price discovery, reproduction, context recovery, correction, retesting, rereview, and documentation…
-
When AI raises PR arrival rate faster than review capacity
AI can increase pull request arrival without increasing safe delivery. Use a queue model to expose review capacity, work in progress, and…
-
The agent picked the wrong files in a large repo
A realistic large-repo failure: the AI edits plausible files, review looks busy, and the relevant module is discovered too late.
-
The documentation was correct when the AI rollout started
A realistic enterprise AI failure: the docs were correct at launch, then the code moved and agents kept following old instructions.
-
New developers ask the AI the same questions every week
A realistic onboarding problem: new developers use AI to ask the same repository questions every week because the project memory is not…