Why Linear Built an API For Agents

Key Concepts

AI Agents, Software Development, Cursor, Linear, Integration, Agent API, Background Agents, Terminal Access, Issue Tracking, Developer Productivity, Context, User Interface, Code Review, Best Practices, Cost Optimization, Multi-Agent Interaction, Future of AI, Email Agents.

Introduction

This podcast episode of "The New Stack Agents" features Andrew Milik from Cursor and Tom from Linear discussing their recent partnership and the current state of building with AI tools in software development. The conversation covers the integration between Cursor and Linear, the agent API, background agents, and the future of AI in coding.

Background of Guests and Companies

• Andrew Milik (Cursor): Head of Product Engineering at Cursor, previously worked on Skiff (a privacy-focused email product sold to Notion) and Notion Mail. He emphasizes product-led growth and the significant productivity gains (100-200%) enabled by Cursor.
• Tom (Linear): Works at Linear since 2021, a company focused on building an excellent issue tracking tool and a platform for building products. He also runs Outline, an open-source team documentation product. Linear aims to replace tools like Jira with a faster and more efficient solution.

Cursor and Linear Integration

• Motivation: Linear aimed to make agents first-class citizens within their platform, allowing users to assign issues to agents like Cursor. Cursor, in turn, sought to integrate its background agent with platforms where tasks are delegated.
• Functionality: The integration allows users to assign issues in Linear to Cursor. Cursor's background agent then works on the task, providing updates and feedback in a dedicated thread within Linear. This includes information about the files being accessed, questions, and progress towards a pull request.
• Agent API: Linear developed an Agent API to streamline interactions with agents. This API provides a dedicated session for agents, abstracting away some of the complexities of the existing GraphQL API.
• Early Results: High demand from existing customers to connect the background agent to Linear. Enterprises are actively exploring this integration.

Cursor's Background Agent

• Core Technology: The coding agent is the core technology, constantly being improved for better capabilities and longer run times.
• Deployment: The agent can run in the cloud (on Cursor's infrastructure), via the CLI, or within the IDE.
• Developer-Centric Approach: The design emphasizes developer involvement and review of the agent's output, rather than fully automated task completion.
• Terminal Access: Cursor's agent now utilizes the terminal, enabling it to execute commands, test APIs, and interact with the program's output. This significantly improves its capabilities (referenced Apple machine learning paper showing a 30% improvement).

Context and Data Ingestion

• Importance of Context: Linear provides rich context to the agent, including stack traces, support tickets, and debugging conversations.
• Prompting: Clear and well-specified prompts are crucial for effective agent performance.
• Team Setup Context: Teams can provide additional context to the agent, such as information about the codebase or specific instructions.
• MCP (Machine-Readable Configuration Protocol): Connecting to an MCP server allows the agent to access documentation and other integrations.

User Interface and Experience

• Linear's Approach: Linear aims to create a user interface that feels like interacting with another team member. They are exploring ways to differentiate the agent experience, such as forking conversations and structured responses.
• Cursor's Approach: Experimenting with a new layout that moves the agent conversation from the right sidebar to a more central position. This allows the agent to direct the user's focus and open files as needed.

Challenges and Areas for Improvement

• Agent Performance: The quality of the code generated by the agent is a key area for improvement.
• User Education: Teaching users how to effectively prompt and interact with agents is essential.
• Context Overload: Managing the amount of context provided to the agent to avoid overwhelming it.
• Setup Cost: Setting up custom environments for background agents can be complex.

Best Practices

• Clear Prompts: Provide well-defined and specific prompts to guide the agent.
• Chunking Work: Break down tasks into smaller, manageable chunks.
• Leveraging Existing Resources: Point the agent to existing pull requests or documentation for guidance.
• Investing in Developer Experience: Optimize development processes (e.g., fast type checks) to enable more iterations.
• Custom Environments: Set up custom base images and environments for background agents.

Cost Optimization

• Cost vs. Value: The cost of running agents is generally considered inexpensive compared to the value they provide in terms of time saved and bug fixes.
• Prioritization: Focus on high-quality agent runs rather than running agents on every task.

Multi-Agent Interaction

• Coordination: Exploring ways for agents to interact and coordinate with each other, such as passing tasks between agents (e.g., Sentry agent to Cursor agent).
• Code Review Agents: Code review agents are seen as particularly valuable for identifying and preventing bugs.

Future of AI in Software Development

• Email Agents: Improving AI-powered email agents for tasks like replying and auto-labeling.
• Home Assistants: Integrating AI into home assistants to improve their capabilities.
• Seamless Integration: Creating a more seamless and integrated experience across different tools and platforms.
• Planning Focus: Exploring the use of agents for planning and scoping out tasks.

Notable Quotes

• Andrew Milik: "For me, it was not like 10, 15, 25% [productivity increase], it was like 100%, 200%... it was a massive change in how I was doing work."
• Tom: "Linear is a place where you kind of decide and plan what work gets done."
• Tom: "The context is what the agents often really need to to do a really good job."
• Andrew Milik: "The results from the cursor agent could get better and better and better."
• Tom: "These things aren't magic. You still need to give it some some idea of of how to do things."

Technical Terms

• AI Agent: A software program that can autonomously perform tasks.
• IDE (Integrated Development Environment): A software application that provides comprehensive facilities to computer programmers for software development.
• CLI (Command-Line Interface): A text-based interface used to interact with a computer system.
• Async Agent: An agent that can run in the background without requiring constant user interaction.
• GraphQL API: An API (Application Programming Interface) that allows clients to request specific data.
• MCP (Machine-Readable Configuration Protocol): A protocol for accessing and managing configuration data.
• Sentry: An error tracking and performance monitoring platform.
• Data Dog: A monitoring and analytics platform for cloud-scale applications.
• Stack Trace: A list of function calls that led to an error.
• Pull Request: A request to merge code changes into a repository.

Conclusion

The integration between Cursor and Linear represents a significant step towards incorporating AI agents into the software development workflow. While challenges remain in terms of agent performance and user education, the potential benefits in terms of productivity and efficiency are substantial. The future of AI in software development involves creating more seamless and integrated experiences, improving agent capabilities, and exploring new use cases such as planning and code review. The speakers emphasize the importance of providing clear context to agents, investing in developer experience, and continuously iterating on the user interface to create a more intuitive and effective workflow.