Cursor’s Founder on Why You Need Agent-First Infrastructure

The real signal for engineering teams isn't another model announcement—it's the deliberate shift from using off-the-shelf LLMs to training models (via RL) that deeply understand your toolchain. Cursor’s composer 1.5 excels at grep, semantic search, and sub-agent orchestration because these behaviors are baked in during training, not bolted on later. For teams building AI-powered tooling, this suggests a future where competitive differentiation comes from models that natively use your specific APIs, repos, and workflows.

The infrastructure story is equally stark: running millions of sandbox environments for RL training forces you to build orchestrators that no cloud provider offers. Long-running cloud agents, which currently feel like a degraded experience, will only become viable when the model can test its own output—and that requires your company’s dev environment to be machine-readable. The implication: invest now in “dev ex for AI”—clear service startup orders, reproducible setups, and agent-friendly documentation—or your agents will silently fail.

Context management is evolving from prompt tricks to trained behaviors. RL incentivizes models to produce genuinely useful summaries and to grep past conversations, meaning multi-hour tasks won’t hit a wall. This makes it plausible to delegate large, open-ended work to agents, but it also changes the review bottleneck. If an agent runs for a day and produces a 1,000-line diff, the model should prove its own correctness, not just dump code. Teams should start planning for verification pipelines that are automated, not human-in-the-loop.

The long arc points to a “managerial” engineering role where you allocate budget and goals rather than craft every line. While the “self-driving codebase” idea is still speculative, the capability jumps are real and accelerating. The pragmatic take: adapt your workflows now—review everything in long-lived code, but let AI handle ephemeral experiments. The art of engineering isn’t dying; it’s becoming more about system design for AI collaboration.