Andrej Karpathy posted a question this week that sounds deceptively simple: what would an IDE actually look like if it were built for an agent rather than a human? The post landed differently than the usual AI discourse because Karpathy wasn't pitching a product or a paper — he was naming a gap that most of the industry has been quietly working around.
Today's leading AI coding tools — Cursor, GitHub Copilot, and their peers — were designed with a human programmer at the center. The AI handles discrete, bounded tasks; the developer reviews, decides, and moves on. That model made sense in 2022. It's starting to look like a constraint. As agents get better at multi-step execution, the human-in-the-loop assumptions baked into every layer of these tools become friction rather than guardrails.
The products pushing hardest toward autonomy already feel the mismatch. Cognition AI's Devin runs multi-step engineering tasks end-to-end; Aider handles a more modest agentic loop in the terminal. But both still sit on top of infrastructure built for a very different workflow — one where errors are surfaced to a human, context resets between sessions, and version control is something a person manages. None of that holds in a genuinely agentic setup.
Karpathy's framing clarifies what building for that setup would actually require: persistent task memory across sessions, sandboxing that lets an agent safely run arbitrary code in an iterative loop, and version control that treats exploratory branches as first-class citizens rather than afterthoughts. The most direct answer emerging from the discussion is a unified workspace — terminal, browser, editor, and test runner collapsed into a single environment designed around the agent's needs rather than assembled from tools designed for ours.
The timing matters. Model capability has been the dominant bottleneck for the past few years, and that story is getting harder to tell as models improve faster than the infrastructure surrounding them. Tooling is where the gap is now, and Karpathy's question — simple as it sounds — is a signal that the next wave of competition in developer tooling won't be about adding AI features to existing IDEs. It'll be about whether anyone builds a new one from scratch.