Codex got the same god class Claude had failed to remove, plus one thing Claude never had: a target architecture. An implementing agent and an independent reviewer then took three rounds to converge, because agents renovate around bad architecture rather than delete it.
A boolean feature flag starts a risky migration everywhere at once. Hashing a stable tenant ID into 100 buckets turns the flag into a dial that only ever widens, and the thresholds matter: percent-of-tenants is not percent-of-work.
Handing an AI your engineering conventions cleans up the code without fixing the architecture. A 5,000-line god class shrank to 500 lines across tidy domain files, and an independent reviewing agent still returned FAIL: authority never left the central object.
Infrastructure documentation goes stale because it is a second thing to maintain. Terraform plus cloud discovery plus a few query tools describes the system better than any diagram, and leaves humans to document only the why behind the odd decisions.
17 lock-wait timeouts in two hours. The SELECT was doing exactly what it was supposed to. The problem was what InnoDB did silently in addition: a gap lock on the write path, invisible in logs, invisible in EXPLAIN, visible only in production.
Most architecture methods try to predict which failures will hit. Residuality Theory does the opposite: it throws random stressors at a design and keeps whatever survives. The randomness is the point, because resilience that depends on predicting the future is not resilience at all.
Auditing data-heavy companies reveals the same pattern: asynchronous data processing crammed into the synchronous web stack. The contention shows in performance, delivery, and team dynamics. Isolation fixes all three.
LLMs are no longer a tab you open. They're the interface layer between intent and every system underneath. This post maps what ambient AI, edge inference, and agent-as-infrastructure mean for how you design modern software.
Every legacy codebase is a palimpsest: layers of decisions written on top of each other, none fully erased. Geoffrey Dhuyvetters argues that what looks like technical debt is often stratigraphy, and you read it before you rewrite it.
A mobile app, past its usefulness, was days from being phased out. One email reversed the decision. No discussion. No input from engineering. This is what happens when decision-making drifts too far from the work.
The machines aren't replacing developers, they're promoting them. You're no longer just writing code; you're managing agents, reviewing output, and setting standards. Three Claudes walk into a codebase, and suddenly you're a manager.
Last week, we deleted 16,000 lines of code, rewrote 11,500, migrated frameworks, ripped out React, and swapped our entire CSS setup with AI. It took less than 6 hours.
Technical debt used to justify meetings, trade-offs, and dedicated sprints. AI has changed that. Cleanup is now fast, cheap, and continuous. Teams that stop debating and start fixing unlock faster delivery and better outcomes.
A couple of months ago, I was copy-pasting prompts into ChatGPT. Now I'm shipping features, running tests, managing branches, and keeping documentation alive, with a team of agents doing the heavy lifting. All by myself.
Software development's feedback loop has compressed from years to minutes, but QA remains the last bottleneck, the one place still dependent on human judgment. AI is rapidly closing that gap, and before the year is out, that final human checkpoint may no longer be necessary.
