Why Linear Built an API For Agents
A few weeks ago, Cursor, the popular AI code editor, and project management tool Linear, which competes with the likes of Atlassian’s Jira, announced an integration that would allow Linear users to bring the Cursor agent right into their project management flows to fix issues, implement design changes and more.
To learn more about this integration, as well as about how to best use these autonomous coding agents in any environment, I talked to Linear’s head of engineering, Tom Moor and Cursor’s head of product engineering, Andrew Milich, for our latest episode of The New Stack Agents.
Neither Moor nor Milich could remember who approached whom to begin this partnership, but since the Cursor team was already using Linear as its tool of choice for tracking tasks and issues, bringing the Cursor agent right into Linear must have felt like a natural fit.
Meanwhile, Linear was also working on its platform to enable agents to become first-class citizens inside of Linear.
With the rise of coding agents running in the background, we’ve seen quite a few similar integrations into tools like GitHub, after all, and since Cursor recently launched its background agent, it had all the necessary pieces in place for this integration.
With the integration, developers can now assign a task in Linear to the Cursor agent and then have the agent work on this in the background.
Best Practices for Background Agents
It’s no secret that we’re still in the early days of working with background agents. Many developers are still building up their intuition for how to best assign issues to an agent, for example.
“People are actually, in the hundreds of thousands or millions, turning these things on and getting a lot of value from them,” Milich said. “And even for a normal Cursor user at Cursor, seeing it ask for GitHub issues, or ask for Linear issues, then open a browser and check Sentry, all with [Model Context Protocol] or in the background, is pretty amazing.”
But, as Moor also noted, a lot of the current work has to be about teaching developers what these tools are capable of. Developers still have to spend some time thinking through the problem themselves to guide the agent.
“It’s not unusual, even on our own team, to see an issue that has just a title, and then someone says: ‘@cursor, fix this,'” Moor said “What did you expect to happen exactly here? Like these things aren’t magic. You still need to give it some idea of how to do things. And I think just spending that one minute doing the thinking, some of the thinking yourself, to give it this head start goes a long way.”
One tip Moor had was to point the agent to an existing pull request where either the agent or a developer did something similar to what the agent should work on now.
Milich added that there are some use cases where things are easier for the agent, especially when it comes to things like updating docs, fixing dependencies, or working with data.
Building An API For Agents
What’s interesting here is that, as Moor told me, Linear actually built out an API specifically for agents to enable this integration.
“For the first version, we just had the agents use our existing GraphQL API, which is very extensive and something we’ve been building for five years now,” Moor said. “You could pretty much clone linear on top of our API without too much trouble.”
But the team also realized that there are a lot of surfaces inside of a tool like Linear that an agent should have access to for features like dealing with comments, replies and comment threads.
“So we kind of rethought it and created this idea of an agent session in our API, which I don’t think is a unique concept. I think most agents have something similar,” Moor said. “So we wanted to codify that, where when a user mentions you, mentions the agent, mentions Cursor, or assigns Cursor, we send Cursor this concept of an agent session with all the context.
“And then Cursor just has to hit our API with the things it’s thinking and its responses with that agent session ID, and it abstracts away the user interface for the most part. You can still use the full GraphQL API to do anything that might make sense to do, though.”
For Cursor, it was the launch of Linear’s agent API that made the team decide to go ahead with the integration.
“I think all of these things were mostly buildable, and we had built a version with the old API,” Moor said. “But then, the first-class agent support with an agent conversation and letting users submit follow-ups and see the agent pane on the side, I think that had a lot of buy-in from our team that we can build something that I think is really good.”
Milich also noted that it’ll be interesting to see if more companies end up building APIs specifically for agents.
For more of our conversation, including Milich’s and Moor’s thoughts about how we’ll interact with these agents in the future, building multi-agent systems, and what they wish AI could do to make their own lives easier, check out the video of our conversation on YouTube or subscribe to our podcast here.
