Cursor's CEO on the agent-first shift, plus a first look at a frontier model trained from scratch
Opening Keynote, Michael Truell | Compile 26 · Michael Truell
27 min total·Actually worth watching closely: ~14 min·3 must-watch clips
- 0:00 – 2:45Listen
The prehistory before the product
Truell looks back at how Cursor started — four young programmers in the prehistoric times of January 2022, the first version shipping at the beginning of 2023, and the detour they took at first because dozens of startups were already working on AI coding and they were nervous they couldn't compete.
They almost didn't build an AI coding tool at all; what got them there was wanting to build something useful, not confidence they would win the category.
Pure spoken company history with nothing to look at — fine for the commute.▶ Jump to 0:00Speaker · Michael Truell - 2:45 – 5:25Listen
The hard numbers on the shift
Agents are now the primary way over 95% of users work, roughly 5x assistive features like Tab by request volume — and many orders of magnitude more once you control for lines of code. The transition took about a year and a half and only settled in the last three months.
By the numbers, the completion era of AI coding is over and the center of gravity has moved to agents.
It's a few percentages and multiples — hearing them is enough, the slide adds little.▶ Jump to 2:45Speaker · Michael Truell - 5:25 – 7:05Watch
Cursor 3 re-lays out the interface
A look at what Cursor looks like now: agents on the left, agents in the center, the whole interface organized around agents rather than line-by-line editing — while keeping file editing, remote SSH and always-on language servers.
The layout change is itself the evidence: the editor steps back, and managing agents becomes the main view.
The persuasive part is entirely on screen — you need to see the before and after side by side.▶ Jump to 5:25Speaker · Michael Truell - 7:05 – 11:00Skim
From closed product to something you can call
The SDK makes Cursor programmable and accessible to other agents; a plugin ecosystem lets agents use the same tools engineers use — Datadog, databases, design tools; and work can be handed off straight from Slack, Jira and Linear.
Cursor wants to be a component inside someone else's agent system, not just an app you open.
Mostly a run-through of integrations and capabilities — skim to see whether your tools are on the list.▶ Jump to 7:05Speaker · Michael Truell - 11:00 – 14:05Listen
Two training tracks behind the in-house models
A recap of the two paradigm shifts the Composer series brought: continued pre-training on tens of trillions of tokens on top of open source models, and real-time RL that updates the model every five hours from online user feedback.
A new model every five hours is the most counterintuitive engineering fact here, and the technical basis for their insistence on speed and cost.
Training methodology and values, no demo — worth listening closely.▶ Jump to 11:00Speaker · Michael Truell - 14:05 – 18:10Watch
Giving agents their own dev environment
Kevin takes the stage on cloud agents: today developers only hand over 30 minutes to an hour of work and run three to five agents locally, bottlenecked by their own machine. The fix is a computer per agent, with cloned repos, installed dependencies and credentials for build systems and internal tool chains.
Handing a whole project to an agent doesn't hinge on a smarter model — it hinges on the agent having a full environment where it can run tests and verify its own work.
The stretch where the agent works autonomously, verifies itself and hands back a screenshot you can circle is the densest visual in the talk.▶ Jump to 14:05Speaker · Kevin - 18:10 – 21:10Watch
Managing agents by phone, and bulk migration in the background
Tomas demos Cursor Mobile (iOS beta, TestFlight open that day) — seeing which agents are running or blocked and annotating what they produce, from anywhere — alongside automations passing 6 million runs in its first months and customer Amplitude migrating 20,000 React component instances to their Tailwind equivalents with background agents.
Usage is moving from prompting line by line to setting up loops and letting agent systems run; the 20,000-component migration is the best sense of scale.
Annotation and remote control on mobile are hard to convey in words and worth watching live; for the customer numbers later, listening and noting the figures is enough.▶ Jump to 18:10Speaker · Tomas Reimers - 21:10 – 26:38Watch
The Origin platform and the next model
Origin launches: Git infrastructure rebuilt on cloud provider primitives, able to take thousands of agents pushing and pulling at once, with automatic conflict resolution, CI fixes and replies to PR comments — said to cut review time by more than half, open to everyone this fall. The talk closes by announcing a frontier-class model trained entirely from scratch, not on an open source base, on 10-20x the compute and aimed beyond coding at general intelligence, shipping in the next couple of weeks.
That closing model announcement is the heaviest piece of news in the talk: Cursor is no longer only working at the application layer, it's entering the frontier model race head-on.
Origin's interface and automatic conflict handling actually show up on screen, and the closing announcement is worth watching along with the visuals.▶ Jump to 21:10Speaker · Michael Truell