Figma deploys internal AI systems to automate experiments and PM tasks
Figma has built internal tools including AutoXP for growth experiments and PMOS for product managers. These systems use AI agents to automate analysis, writing, and task management, reducing human intervention in routine workflows.
Figma product manager Ezra McCaber developed an internal automation system called AutoXP to maximize experimentation cadence. The system operates via three Cursor automations triggered by Slack messages, orchestrating a workflow involving a data scientist agent, a UX writing agent, and an engineer agent. It communicates in a dedicated Slack channel, writing back to the original thread to track progress. Currently, an engineer named Marianne is running five concurrent live experiments on surfaces the team would not normally prioritize. The system found a significant win on its very first experiment. The workflow requires human review only before launch to validate data science findings, hypotheses, and code; after launch, the system automatically triggers result analysis.
Figma has also developed an internal repository called PMOS, containing 31 skills designed to assist product managers. The system originated from a challenge by former CPO Yuki to make PMs more effective with AI. It was built by McCaber, Jenna, and Lawrence, who refined the skills during an offsite with all Figma PMs. PMOS is structured as a Git repo with an agents.md file, skill definitions, and command hooks for version synchronization across 60 to 70 PMs. Key skills include 'What's Up,' which aggregates Slack and project data to highlight outstanding decisions, and 'Build Artifact,' which generates HTML dashboards. McCaber now uses PMOS primarily through Codex, leveraging its browser-use capabilities for tasks like bug bashing.
McCaber also built a prototype real-time writing tool using the Jev model, a fast and cheap AI model. The tool, built on the 'Marker' writing app, integrates with Figma's internal data library. As the user types, the system triggers a Jev check whenever typing pauses, scanning for incomplete thoughts or claims requiring substantiation. If a gap is detected, the system queries the internal library for relevant data and suggests completions in real time. McCaber argues that Jev's speed and low cost make it uniquely suited for real-time interactions, where top-tier models are too slow to feel like a collaborative coworker.