Figma describes Nimbus, a centralized platform managing tens of thousands of sandboxes for untrusted AI-generated code across multiple cloud providers. It covers namespace-based APIs, read-only outbound network policies, Postgres-based warm pool claiming, and health-driven provider routing. Reported results include a 90% latency cut and availability rising from 99% to 99.95%.
Figma engineering blog
Loading
Figma's engineering team describes training a first-party design editing model, starting with supervised fine-tuning on synthetic edit data and teacher-model validation, then moving toward reinforcement learning to improve generalization. The post also covers how they evaluate design edits across separate structural and visual dimensions, and reports speed, undo-rate, and preference results from the Figma agent's rollout.
Coinbase’s Design System team evaluated Figma Code Connect with coding agents and found that richer component context improved design-system adherence while reducing token usage, implementation time, and cost. The case study details the evaluation design, workflow, and measured gains, including a 22.5% average cost reduction.
Riccardo Erra demonstrates a workflow for keeping AI-generated code aligned with design intent using Figma MCP, FigJam diagrams, Code Connect, and design-system search. The process covers planning interactions, reusing mapped components, detecting token drift, and synchronizing new UI states before review and release.
The post appears to explain how Figma structures engineering critiques to make feedback more effective and improve technical collaboration. The body was unavailable, so this summary is based on the title alone.
The post discusses how engineering is changing in the age of AI, but its body was unavailable, so no specific technical practices, evidence, or takeaways could be verified.
The post appears to discuss how Figma built its code layers, but the body was unavailable, so its specific technical approach and takeaways could not be verified.
Figma introduces an MCP server intended to bring Figma capabilities into developer workflows. The post body was unavailable, so the summary is based on the title and URL alone.
The post discusses how Figma’s Code Connect works with the Model Context Protocol (MCP) to produce better code while using fewer tokens. The body was unavailable, so implementation details and supporting evidence could not be verified.
Figma’s Data Science team explains how it used a randomized controlled trial to measure Figma Make’s impact while controlling for task complexity and participant differences. The study found a 20% reduction in task-completion time overall, with product managers seeing the largest gains.
Figma’s security team built an agentic alert-response system that combines RAG, scoped tools, Snowflake investigations, procedural memory, and deterministic guardrails. The system reduced complex-alert resolution time by about 70% and on-call pages by 20%, while generating reviewable remediation pull requests.
Figma details how it uses AI agents to prevent, detect, and fix vulnerabilities across code generation, pull-request review, and decade-old monorepo audits. The team explains its shared threat-model policy, precision and recall evaluation, multi-model review, adjudication, feedback loops, sandboxing, and measured security outcomes.
Figma explains how it rebuilt browser accessibility semantics for a canvas-based editor using an internal accessibility tree, a synchronized Mirror DOM, and bidirectional focus and selection handling. The post also covers spatial positioning, ARIA live-region announcements, and incremental updates for responsive screen-reader interactions.
Figma explains how it rebuilt the Layers panel using two-pass computation, windowing, lazy derived properties, and rope-like structural sharing. The redesign reduced unnecessary work, cut cache memory growth to O(n), and made key interactions 30–50% faster in large files.
Figma explains how it replaced PgBouncer with PGKeeper, a Go and gRPC service for PostgreSQL connection management, admission control, and load shaping. The post details prioritized backpressure, CoDel-based queue handling, fair sharing, connection-churn avoidance, and the staged rollout that protected database reliability at scale.
Figma details the design and rollout of FigCache, a stateless RESP proxy and first-party client ecosystem that decouples Redis connection scaling from application fleets. The platform centralizes routing, observability, failover handling, and extensibility, achieving six nines of caching uptime and an order-of-magnitude reduction in Redis connections.
Figma explains how it replaced its decade-old Instance Updater with Materializer, a generic reactive framework for maintaining derived document subtrees. The push-based invalidation model, automatic dependency tracking, shared runtime orchestration, and side-by-side production validation delivered 40–50% faster updates in large design files while enabling new dynamic features.
Madison Kohls explains how data scientists can create impact in complex, correctness-critical domains such as billing. She details building an invoice-reconstruction application and SQL-based consistency checkers that improve explainability, detect state and data-quality errors, and provide observability across systems.
Figma describes Response Sampling, a lightweight asynchronous middleware system that samples API responses, validates file-identifier permissions, and detects sensitive fields marked by its FigTag data-classification system. The post explains its staging and production architecture, false-positive controls, performance trade-offs, and defense-in-depth benefits.
Figma explains its migration from WebGL to WebGPU, covering graphics-interface redesign, shader translation, uniform-buffer batching, and shared WebAssembly and native implementations. The post also details performance benchmarking, device compatibility testing, and dynamic fallback to WebGL during rollout.