Cursor's VP of Engineering on Building AI Agent Teams

In a recent 'Cursor Conversations: Behind the Build' session, Tido Carriero, VP of Engineering at Cursor, shared insights into the evolving role of AI in software development and the critical challenges of building and managing AI agent teams. Carriero highlighted the dramatic increase in AI-generated code, noting that approximately 60% of enterprise merged commits are now written by agents, a figure that has seen exponential growth.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with funding and a one-liner from our database.

Cursor

$50.0B

An AI-native code editor that helps developers write, understand, and manage code more efficiently.

Linear

$134M

Fast, elegant issue tracking and project management system for software teams and AI agents

Statsig

$70M

Feature flagging and product experimentation platform for engineering teams.

The agent-driven SDLC

Carriero outlined a vision for an 'agent-driven SDLC' comprising four key phases: Plan, Build, Ship, and Retro. He emphasized that while AI agents are becoming highly proficient in tasks like code generation and architectural explanation, human involvement remains essential. The current challenge, he noted, is to identify which parts of the process humans should still handle, such as reviewing product plans, architectural decisions, and providing crucial feedback to the AI agents.

The Role of Humans in the AI Era

Carriero illustrated this with examples from Cursor's own development process. He described how the company is leveraging AI agents for tasks like triaging issues and identifying security vulnerabilities. However, he stressed the importance of human oversight, particularly in the 'plan' and 'review' stages. For instance, a Product Manager (PM) agent might triage incoming issues, but a human PM is still needed to refine the plans and ensure they align with broader business goals.

Similarly, an Engineering Manager (EM) agent can loop in the relevant engineers for specific tasks, but human judgment is vital for understanding the context and potential implications of changes. Carriero shared an anecdote about a bug report that was initially flagged by an agent but turned out to be a feature request, highlighting the need for human discernment.

Automating Growth Experiments

The conversation also touched on growth experimentation, where AI agents are being used to automate aspects of the experiment lifecycle. Carriero explained how they are synchronizing their roadmap with experiments, using tools like Statsig for A/B testing and Linear for issue tracking. He noted that automating tasks like auditing setup, syncing documentation, monitoring experiment runs, and deciding on the 'winner' of an experiment can significantly increase efficiency.

Security Bots and Automation

Carriero also highlighted the development of specialized agents, such as a 'Security Bot' for auto-patching vulnerabilities. This bot can analyze pull requests, identify potential risks (critical, high, medium, low), and even suggest fixes. He also mentioned an 'Auto Approver Bot' that can automatically approve low-risk commits, freeing up human reviewers for more complex issues. These automations, he explained, are not just about efficiency but also about creating a more robust and secure development process.

The overarching theme was the symbiotic relationship between humans and AI in modern software development. While AI agents can handle repetitive and data-intensive tasks, human expertise, creativity, and critical thinking remain indispensable for driving innovation and ensuring the quality and security of software.