How Cursor builds agentic workflows across the SDLC

By Cursor

Categories: AI, Tools

Summary

AI-assisted coding has plateaued at 40% productivity gains despite handling 98% of merged commits at forward-leaning companies—the real bottleneck isn't code generation but automating the entire SDLC (planning, shipping, validation, retros), requiring agents to handle artifacts like architecture decisions, demos, and production approvals.

Key Takeaways

  1. Coding assistants hit a fundamental productivity ceiling around 40% improvement; jumping beyond requires automating the full SDLC, not just the build phase. Expected 3-4X gains from collapsing 60-80% of R&D spend on handcrafting code, but most teams plateau earlier.
  2. At Cursor, 98% of merged commits are AI-generated—but the company still requires humans at five critical gates: plan approval (before investment), architecture review (key decisions and risky code), production sign-off, incident response (PagerDuty), and outcome feedback loops.
  3. Agents must produce human-readable artifacts at each SDLC stage: rigorous plans in planning phase, architecture explanations and demos in build phase, and monitored deployments in shipping/retro. Visual artifacts (screenshots, demos) are non-negotiable for PR review.
  4. The bottleneck shifted from code generation to orchestrating agent workflows across planning, building, shipping, and retros. Success requires removing downstream SDLC bottlenecks, not optimizing the coding phase alone.
  5. Enterprise AI code adoption jumped from 0% (early 2025) to 15-20% across enterprises, with 75% mentioned as a more recent benchmark. Forward-leaning companies are already at 98% AI-generated commits, signaling a leadership/laggard split in adoption velocity.

Topics

Transcript Excerpt

So I'm going to talk a little bit about what's going on behind the build here. So give you a peek into some of the things we're doing as we enter this third era of building agent teams. Most of this will be pretty tactical and just things we're actually doing since I find it's a very hard topic to get into and figure out how you actually start building these teams, and so I thought it'd be useful to share notes from the field. But before I do, I wanted to motivate a little bit with sort of as I've been thinking about building agent teams, what are the hard challenges for an engineering leader, because I thought it would be good to say these things out loud. So what are the new challenges? Well, the new challenges are, as Michael alluded to, we've gone from zero amount of code being written…