I wrote my last significant function by hand in the spring of 2025. Not the last function I shipped—the last one I typed. Since then every line that reached production at Propstreet and Loggad passed through an agent first, and my hands moved from the keyboard to the harness: the prompts, the tools, the gates, the evals that decide whether the agent’s work ships.
This is not a story about typing less. Authors and editors read the same manuscript differently. The author asks: is this what I meant to say? The editor asks: is this true, is it clear, and will it survive contact with the reader? When the agent writes the code, you stop asking the author’s question. You have to. There is too much of it, and it arrives too fast.
The editor’s job
The editor’s tools are not the author’s tools. Mine are a review queue, a test suite the agent must satisfy but cannot edit, and a set of evals that encode what “good” means for this codebase. The agent proposes; the harness disposes. On a normal working day the exchange looks like this:
> implement the retry policy from #482, evals must stay green
14 files changed · 3 tools gated · approval requested: schema migration
> approved. run the eval suite against the prod snapshot
41/41 suites passed · cost $0.87 · trace 4f2a…Notice what I did there and what I didn’t. I set the goal, named the constraint, granted one permission, and read one summary. I did not read fourteen files. The suite read them for me, the way a fact-checker reads for an editor. Trust in the process replaces attention to the artifact—but only if you built the process, and only if the process can say no.
Judgment through systems
The uncomfortable part is status. Authors get bylines; editors get mastheads. If your sense of craft lives in the shape of the code you personally produced, this transition will feel like loss, and pretending otherwise would be dishonest. Mine did, for about a month. Then a release shipped in a week that would have taken a quarter, with better test coverage than I would have written by hand, and the feeling resolved itself.
The rest of this book is the harness in detail—tools, gates, budgets, evals, traces—because the editor is only as good as the desk. This chapter’s claim is smaller and comes first: the skill that survives is judgment, exercised through systems, and the developer who accepts the editor’s chair early will out-ship the author for as long as this era lasts.