The Self-Driving IDE Is Coming

One of the few essays Steve wrote from inside the machinery he's describing: in May 2024 he was personally wiring Cody's inline-edits feature into JetBrains, and the pain is fresh on the page — dozens of new protocol types, round-trips "requiring a sequence diagram to fully explain," and the observation that every IDE plugin has "exactly one zillion listeners." He promises a top-down telling with no stories or jokes, and holds out for roughly four paragraphs before the analogies escape: The Truman Show, a cloned home-security system you fool by dressing your family up as Bob's, and a protocol he first "thought was completely insane" until months of hands-on work convinced him "it's actually completely insane." Under the comedy is a plain industrial-economics question: why does every coding-assistant vendor abandon the features-times-IDEs coverage matrix, and what happens if one company pays full price for it?

The page files this as a prediction, and it split. The self-driving part arrived fast: within a couple of years AI really was operating developers' tools end to end, well past the demos Steve needles in the late section on agent startups (the section title alone earns the scroll). But the harness that won was mostly the terminal, and the cross-IDE matrix stayed unpaid-for, Sourcegraph's own bet included. The core claim, that agents are only as good as their plumbing into the tools developers actually use, held up much better than the specific plumbing did. Also in here: a personal Emacs scheme, with Cody writing emacs-lisp on demand so Steve can skip his usual six hours of Info pages — plus an open invitation to come work on plugins, extended to all five people he figures might want to.