The Future of Coding Agents

An encore recorded while the applause was still going: fifty pull requests had landed in the three days since launch, most of the audience hadn't even come back from the holidays, and the confidence is set accordingly. The backstory delivers more embarrassment than triumph. Orchestrator #1 was built on Temporal (the "Bagger 288 of workflow orchestrators," too much machine for the job); orchestrator #2 was supposed to live in a private repo until, Steve suspects, it quietly made itself public; and the Go rewrite needed three or four redesigns because the agents kept guessing wrong about the directory structure. The diagnosis of why the first two failed comes down to a single clean distinction, and it's worth the whole section.

The predictions are the checkable kind, with dates attached: what agent vendors ship in 2026, when open-source models get good enough for garage hackers again (the FAQ pegs them at roughly seven months behind frontier), how badly team coordination breaks at full agent speed. That last one arrives as lunch gossip rather than argument — two ex-Amazon friends moving so fast that a teammate citing two-hour-old information gets told it's ancient, and one big company that met the merge problem by decreeing one engineer per repo. Readers of the models-have-plateaued school get no quarter here, and no evidence either; the essay simply asserts they'll keep improving and moves on. It ends with a song, and the tune is the pun the tool's name has been begging for all along.