Coding After Code Agents
Since Codex and Claude Code, the unit of work stopped being the line of code and became the task. The workflow change is real, and the products show it.
Something flipped after Codex and Claude Code matured. I stopped opening an editor first and started opening a conversation with my codebase. The unit of work is no longer the line of code-it's the task. Describe what must be true when it's done, set the boundaries, and review what comes back.
That shift restructured my whole day. Specs became the primary artifact, because agent output tracks the quality of the brief. Repo convention files turned tribal knowledge into something every session actually reads. Tasks got PR-shaped, since bounded and verifiable is where agents shine. And review became the job: my role moved from writing most lines to judging all of them.
The product side is where it stopped feeling like hype. I maintain several products and their sites in parallel, and the perpetual-neglect tax is gone-migrations and restructures happen the day they're noticed instead of queuing for weeks. This site's entire content layer moved from static JSON to a Postgres-backed CMS with zero downtime in about a day of supervised agent work. A year ago that was a two-week project I kept postponing. The activation energy for architecture improvements dropped below the threshold where procrastination wins.
The catch: gains follow the guardrails. Type checks after every change, tests as the contract, real browser verification, no silent shortcuts. An agent plus a strict reviewer compounds; an agent plus blind trust degrades. The craft didn't disappear-it moved up a level. The products are better for it.