Editor | Da Shi
The opening keynote of Cursor's first Compile conference is here!
Founder Michael Truell took the stage and immediately unveiled three major announcements: Cursor Mobile, Origin, and the next-generation in-house model.
Cursor Mobile is responsible for bringing running Agents to your phone. Origin aims to push Agents into the PR, CI, and review pipeline. The next-generation in-house model is designed to bolster the underlying capabilities for Agents executing long-running tasks.
Truell: Currently, "over 95%" of Cursor's usage comes primarily from Agents. In on-demand invocation, the frequency of Agent/smart assistant usage is about "five times" that of auxiliary capabilities like Tab.
This indicates that Cursor has already shifted its main battlefield from code completion to task execution.
Cursor placed Agent at the core of its product line in the keynote
Cursor Mobile: Bringing Agents to Your Phone First
Cursor Mobile is easily misunderstood as a mobile version of an IDE.
But judging from the conference demo, it looks more like an Agent task dashboard.
Kevin Niparko: Developers can view all running Agents on their phones, see which Agents are blocked, connect to repositories, view artifacts generated by Agents, and see short, verifiable screenshots. If a page style looks off, developers can annotate directly on the screenshot and let the Agent continue making modifications.
Cursor Mobile demo showing how developers manage Agents on mobile
This product detail clarifies Cursor's update objectives.
The phone isn't for writing complex modules; it's for acceptance, annotation, and progress tracking.
Agents can work continuously in the background, and developers don't need to stay glued to their computers. Glancing at task progress on the go, circling issues, and then handing them back to the Agent to continue running—that's the scenario Cursor Mobile wants to enable.
In other words, Cursor Mobile's selling point is mobile management of Agents; mobile coding is just the most superficial understanding.
Origin: The Most Underrated Agent-Native Git Platform
The second announcement is Origin.
Tomas Reimers, Graphite Co-founder: Software development "goes far beyond writing code." It also includes testing, reviewing, merging, and deploying.
In the past, AI programming tools mainly solved the problem of writing code. But once Agents start writing code in bulk, the subsequent PRs, CI, reviews, and merge conflicts become new problems.
Cursor defines Origin as an "agent-native Git platform." If an Agent can write code, it will inevitably need to branch out, submit PRs, fix CI issues, and handle review comments.
Tomas Reimers announces Origin, an agent-native Git platform
Origin can handle merge conflicts, fix CI failures, process user comments, and automatically calculate the next steps for each PR, alerting developers only when necessary.
The development team also conducted large-scale tests on Origin: simulating thousands of Agents simultaneously pushing to and pulling from a single repository.
Cursor not only enables Agents to write code but also handles the subsequent hassles: branches, PRs, CI, reviews, and merges.
These are the most time-consuming aspects of software engineering.
Cursor's In-House Large Model: 10-20x Compute Power, Transcending Mere Coding
The third announcement is the next-generation in-house model.
Michael Truell mentioned that Cursor is training a new model fundamentally different from previous ones.
He offered three pieces of information: the model scale is comparable to Opus and GPT; it is "trained from scratch"; and the computational resources used are "ten to twenty times" that of previous efforts.
The most crucial statement is that the new model will "transcend mere coding."
Truell introduces the new model trained from scratch
If an Agent were merely filling in a few lines of code, external models would suffice.
But what Cursor now demands is complete task execution.
Planning tasks, calling tools, reading repositories, running tests, viewing screenshots, judging failures, and retrying.
Stacked together, these capabilities turn the model into the product's linchpin.
Relying solely on external models means ceding control over speed, cost, context, tool invocation, and long-term planning. For Cursor to become an Agent development platform, it must reclaim these underlying capabilities.
A few days ago, TechCrunch, WSJ, and others reported that SpaceX has agreed to acquire Cursor's parent company, Anysphere, in a $60 billion all-stock deal.
This indicates that AI programming tools have been placed within the valuation framework of next-generation development infrastructure.
For developers, whether Musk buys Cursor isn't the key issue.
What matters is: Why is Cursor worth this price?
The answer lies within this keynote: Agents, cloud environments, the Git platform, and the in-house model—all must be controlled in-house.
95% Usage from Agent Explains the Three Announcements
Truell: Over the past few months, Cursor's direction has become clearer: the product needs to be simpler, stronger, and fundamentally agent-first from the ground up.
Data shows that currently, 95% of usage comes primarily from Agents. This proportion signifies a fundamental shift in how Cursor is used.
Users are no longer just asking it to complete code; they are beginning to hand entire tasks over to Agents.
In the past, opening Cursor felt like having faster completions, better context understanding, and smoother code modifications. In that logic, the human was still in charge, with AI playing a supporting role.
Truell mentions that over 95% of Cursor usage comes primarily from Agents
This data set explains why Cursor simultaneously released Mobile, Origin, and an in-house model.
Cursor Mobile addresses how humans manage Agents anytime, anywhere.
Origin addresses how Agents enter team collaboration workflows after writing code.
The in-house model addresses whether Agents can handle longer, more complex tasks.
These three announcements, spanning mobile, Git platform, and model infrastructure, all revolve around a single question: when Cursor's core usage comes from Agents, how should development tools be rebuilt?
Agents Need Their Own Computers Too
Throughout this conference, Cursor repeatedly mentioned giving "Agents their own computers."
Long-running Agent tasks can't always depend on a developer's local machine. They need to clone repositories, install dependencies, authenticate access to build systems and internal toolchains, run tests, generate screenshots and artifacts, and continue retrying after failures.
This already surpasses the capabilities of ordinary editor plugins.
Kevin Niparko, in the Cloud Agents segment, noted: Once an agent system is running, it might work continuously for hours, days, or even weeks. If it's to handle the heavy lifting in the development process, it must always be online and available.
He also gave two figures: over the past few months, cloud agent speed has increased threefold, with reliability reaching 99.9%; since the automation product launched, it has completed over 6 million automated runs.
Cursor positions Cloud Agents as a piece of development infrastructure
The case of Amplitude illustrates this point. It is migrating 20,000 React component instances, replacing them with corresponding Tailwind components, powered by a custom migration Agent built on Cursor.
This type of work isn't suited to being perpetually tethered to a personal computer. It requires batch modifications, repeated testing, continuous execution, and the ability to continue processing after a failure. The value of Cloud Agents lies precisely here: moving background execution, long-term verification, and batch migrations from a local machine to the cloud.
The Developer's New Role: From Writing Code to Orchestrating Small Agent Teams
Truell: Let Agents complete entire feature developments, entire bug fixes, even entire projects.
A year ago, this might have sounded far-fetched. Today, it's part of Cursor's product roadmap.
For developers, the change isn't that hard to grasp. You'll increasingly break down tasks and delegate them to Agents. How to define an issue, which files are off-limits, how to run tests, what the acceptance criteria are, whether to roll back or continue after an error—these questions will become increasingly specific.
Many underestimate AI programming right here.
An AI modifying a single function versus an AI completing a dependency upgrade, a component migration, or fixing a CI failure—the difficulty is on an entirely different level. A developer's value will increasingly manifest in task decomposition, boundary setting, and outcome assessment.
Final Thoughts
Following this Cursor conference, some netizens praised it effusively:
AI programming tools are evolving from code assistants into development operating systems.
Going forward, what developers face will expand from a chat box into a team of Agents capable of running tasks, reading repositories, opening PRs, and fixing CI.
A programmer's value won't disappear, but will shift from personally writing every line of code to orchestrating a group of Agents to get things done right.