Scaling AI Agents at Figma: Verification and Cultural Guardrails

Prioritizing Deterministic Verification

Figma treats AI adoption as a three-act process, moving from initial experimentation to failure-prone scaling, and finally to a mature state defined by robust guardrails. The most effective strategy for maintaining codebase quality is shifting verification left. Instead of relying on human review for everything, teams encode architectural standards and logic into deterministic flows like linting, compiler checks, and unit tests. When an agent identifies a useful pattern, that logic is codified into a repeatable test rather than left as a prompt-based suggestion. Adopting a Test-Driven Development (TDD) style—where the agent writes code to satisfy a pre-defined test—consistently yields higher quality results than prompting for code and attempting to verify it after the fact.

Planning Over Prompting

To restore developer agency and prevent the burnout associated with endless prompt-response cycles, engineers are encouraged to spend time writing detailed implementation plans rather than just prompting for code. A high-quality plan includes an executive summary (the "why") to prevent agent drift, and it breaks the work into small, independently verifiable phases. Each phase must have clear exception criteria or validation gates. This approach allows engineers to handle the high-level design and decision-making while delegating the implementation to agents. This workflow has enabled significant productivity gains, such as compressing six weeks of manual coding into one week of planning and overnight agent execution.

Cultural Guardrails and Communication

Adoption is often slowest among senior engineers who hold the most institutional context, as they are the first to notice when agents fail. Rather than pressuring them to adopt, teams should hand these skeptics the roadmap for making agents safe; their feedback identifies the exact gaps in verification. Furthermore, because AI has increased the volume of internal documentation and Slack traffic, Figma enforces "attention-aware communication." Every PR description or AI-generated analysis must start with a human-written summary. This convention signals to the reader exactly where to focus their limited attention and distinguishes between human intent and AI-generated content, preventing the perception of "slop" and maintaining trust across the organization.