Becoming an AI Team

Replacing a widespread legacy process across a codebase used to take a dedicated team at Pinterest up to a year. Using modern AI models, they now execute and verify the exact same refactor in under a week. 👇

Pinterest operates infrastructure serving billions of Pins, boards, ads, and real-time signals. Traditional playbooks couldn't keep up with that scale, leaving engineers stuck in constant firefighting. Essential platform maintenance and complex infrastructure updates sat in backlogs indefinitely because manual execution consumed too many human hours.

Pinterest transformed their engineering organization by embedding AI agents directly into operational workflows. Instead of treating AI as an optional assistant, they turned agents into first-class actors that handle heavy execution, freeing engineers to act as strategic orchestrators.

The shift relies on specialized agents taking over specific operational workflows step by step.

When a database cluster breaks down, an AI agent automatically digs through logs, formulates candidate hypotheses, and drafts remediation commands before an engineer ever steps in to troubleshoot.

For data pipelines, platform agents write and validate Spark jobs, allowing product teams to explore complex datasets without needing deep query expertise.

For storage systems like MySQL and TiDB, generative models execute sweeping code transformations, rewriting legacy patterns and verifying the updates.

Engineers move up the stack. Rather than writing boilerplate code or manual scripts, they define system boundaries, write strategic prompts, and validate agent output.

Managers also shift focus away from routine status updates and task tracking. Their primary job becomes filtering out tool noise, establishing psychological safety for fast experimentation, and guiding engineers toward high-impact architectural decisions.

While early AI adoption yielded a modest 20% capacity increase, scaling the model created a surplus of engineering time. Refactoring projects on storage platforms dropped from quarters to days. Years of deferred infrastructure maintenance are now routinely cleared from backlogs, allowing engineers to dedicate their energy to roadmap strategy rather than chronic firefighting.