Articles
Notes on automating software work with AI agents: where the bottlenecks are, how to review what agents write, and what changes for teams. Every article is read and checked by me before it goes up.
Start hereCode review is the new bottleneck: why AI hasn't sped up shippingWhy didn't AI make teams ship faster? Because writing code was never the slow part. Faros AI found high-AI teams merge 98% more pull requests while review time per PR rose 91% and PRs grew 154%. The work piled up in front of human reviewers, so the fix is to change review, not to type faster.Read the article
Bottlenecks
Why faster code did not mean faster delivery, and where the hours actually go.
Code review is the new bottleneck: why AI hasn't sped up shipping
Teams using AI heavily merge 98% more pull requests, but review time per PR grew 91%. Why review became the constraint, and what actually moves it.What the AI coding productivity studies actually measured
One study says AI makes developers 55.8% faster, another says 19% slower. Both are right about different things. A guide to what each study measured.Where the hours go: mapping human wait time in a delivery pipeline
Most of a feature's lead time is waiting for people, not work. How to map wait time in your delivery pipeline and decide which human gates to automate.Amdahl's law for software teams: why faster coding barely helps
If coding is 16% of a developer's time, making it 10 times faster speeds delivery by about 17%. The maths behind the AI productivity paradox, and the fix.
Review
Automating code review, from a second opinion to no person in the loop.
Can you take people out of code review? A staged path to automation
A five-level path from humans reading every pull request to review with no person in the loop, with the conditions that must hold before each step.The writer can't be the reviewer: why a second AI model catches more
Why a coding agent reviewing its own work fails, what correlated errors and self-preference mean for multi-agent workflows, and how to set up a reviewer.Claude Code writes, Codex reviews: what an independent review loop changed
The same game-rebuild task given to Claude Code alone and to a two-model pipeline. 3.7 hours became 1.5, and 389 M tokens became 34 M. Here's why."All tests pass" is a claim. Make your coding agent show evidence
Coding agents report success they haven't earned, and some edit the tests. How to build a review pipeline that trusts exit codes the harness ran, not summaries.AI-written tests that catch bugs, and the ones that only pass
Coverage tells you which code ran, not which bugs a test would catch. How Meta and Google filter generated tests, and a pipeline for AI tests that find bugs.LLM-as-a-judge for code: when it works and when it lies
What the research says about LLM judges: where they match human agreement, where position, length and self-bias distort them, and how to judge code safely.How to reduce code review time when AI agents write the code
AI agents write bigger pull requests: Faros measured +154% in size. Why review maths still favours 200–400 lines, and how to make agents ship small diffs.
Trust
Safety, quality, and how much autonomy to give an agent.
Why developers don't trust AI code, and why they're right to
Trust in AI coding tools fell while use kept rising. Why the distrust is rational, and how to build trust from evidence instead of from the model.Guardrails for AI coding agents: permissions, sandboxes, budgets, stop rules
How to give a coding agent autonomy without the risk: permissions, sandboxes, budgets, stop rules and a trust ladder built on reversibility and blast radius.AI code gets copied, not refactored. What GitClear's data shows
GitClear's data shows refactoring falling and copy-paste rising as AI writes more code. Why agents duplicate, what it costs, and how to make them reuse.Hallucinated APIs and packages: how they get into code and how to stop them
AI models invent packages and APIs. 19.7% of packages in a USENIX study didn't exist. Real examples, why it happens, and the checks that stop slopsquatting.Is AI-generated code safe? A threat model for teams that ship it
AI code fails security tests in 45% of Veracode's tasks. A practical threat model: which risks are likely, which are severe, and the controls for each.
Agents in practice
Running coding agents day to day: specs, hooks, costs, overnight runs.
Claude Code vs Codex: why I stopped choosing and use both
Claude Code and Codex compared on hooks, sandboxing, headless runs, review and billing, and why the best setup uses one to write and the other to review.Claude Code hooks examples: turning an assistant into a pipeline
Practical Claude Code hooks examples: block risky commands, gate Stop on tests, call a second model to review, and the hook mistakes that break a pipeline."Would you like me to continue?" Why coding agents stop halfway
Coding agents finish a third of the task and ask permission to go on. Why that happens, and how stop hooks, checkable criteria and a reviewer end it.The true cost of AI coding: tokens are the small line on the bill
What AI coding really costs: subscriptions and tokens, plus review time, rework and long agent runs. A cost model and how to cut the parts that matter.Spec-driven development with Claude Code: write intent before code
Spec-driven development gives AI agents a spec, a plan and checkable criteria before any code. How to do it with Claude Code and what a good spec contains.Vibe coding vs agentic engineering: same models, different results
Vibe coding and agentic engineering use the same AI models. The difference is specs, checks and review. When each one fits, and how to move up.Running coding agents overnight: what must be true before you sleep
An overnight AI agent can turn eight idle hours into finished work, or into a mess and a bill. The checklist for unattended coding agents that actually works.Why AI agents fail in production: nine failure modes and their fixes
The demo worked, production didn't. Nine ways AI agents fail once they meet real work, what causes each one, and the fix that actually holds up.AI coding agents benchmark: why you should run your own evaluation
SWE-bench scores won't tell you which AI coding agent fits your codebase. How to build a small evaluation from your own tickets, and what to measure.Acceptance criteria an AI agent can't talk its way around
How to write acceptance criteria for AI coding agents: observable, executable, owned outside the agent, with an honest way out. Bad and good examples included.
Teams
How teams, roles and hiring change when agents write the code.
Will AI replace software engineers? The work that moves, and what grows
What 2026 labour data says about AI and software jobs: BLS projections, Indeed postings, Stanford's early-career findings, and which engineering work grows.The dark software factory: levels of autonomy, autocomplete to lights-out
What a dark factory in software development means, Dan Shapiro's six levels of AI coding autonomy, and what has to be true to move from one level to the next.What an AI automation engineer does, and how to hire one
A buyer's guide to hiring an AI automation engineer: what the role covers, what an engagement looks like, interview questions that work, and red flags.From nine developers to one: a product team when agents write the code
A reasoned model of how a nine-person product team can shrink to one engineer on that product while eight move to new ones, and what has to be true first.
Apps
Building Apple apps with agents and with on-device models.