Summary
Eyal Blum, software engineer at Figma, shares an honest account of how Figma’s engineering organization is navigating the uneven adoption of AI coding agents — and what they’ve learned about maintaining code quality as agentic workflows become standard practice.
Blum describes a three-act adoption arc that plays out at both individual and organizational levels: early wins with simple tasks, failure and trust breakdown when applying the same techniques to complex problems, and finally the development of real skill through structured planning, guardrails, and context management. The challenge at Figma is that teams at all three stages must still ship together, creating friction when AI-forward teams work alongside those still in the second act.
Key lessons include investing heavily in verification infrastructure — using Playwright and MCP to enable agents to explore code rather than requiring humans to do it — and writing detailed upfront plans before engaging agents at all. Blum’s planning framework mirrors a design doc: an executive summary explaining the “why” (to prevent agent drift), phases that can each be validated independently, and explicit exception criteria per phase. He also surfaces underappreciated organizational costs: senior engineers becoming bottlenecks as they catch agent mistakes with their hard-to-document institutional knowledge, developer satisfaction eroding in pure prompt-cycle workflows, and AI-generated Slack messages and emails multiplying in volume without increasing in signal. Figma’s approach to restoring developer agency — treating plan-writing as the new craft — is a practical counterpoint to pure vibe-coding workflows.
📺 Source: AI Engineer · Published August 28, 2026
🏷️ Format: Workflow Case Study
