Firebender's AI Coding Agent - Reverse engineered
summary
Firebender is an Android Studio plugin that takes a Figma URL, pulls the preview screenshot, scans the repo for color.kt and theme.kt, then iterates with LLM-driven tool calls (find usages, editor edits, rendered preview diffs) to converge on a Compose screen; it handles ~10 billion tokens/day, uses ~7–9 tools, and two users have accepted over 10 million lines of code each. Engineering choices include Figma preview comparison, speculative file rewrites instead of git diffs, and careful prompt prefix KV cache management; built by three people and used by thousands including enterprises.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:48What Firebender does
- 6:07Real-world impact and usage
- 9:44Processing user requests
- 15:26Prompt caching importance
- 17:55Token throughput and infrastructure
- 18:17Tool overload challenges
- 21:53MCP servers and behaviors
- 23:23Tool configuration risks
- 24:26Evaluating agent performance
- 28:22Cotlin Bench and progress
- 30:10Designing agent evaluations
- 32:14Code editing challenges
- 33:37Firebender journey
related talks
34mMulti-coding agent workflows
34mAutomating Full-Screen TalkBack for a Government App with Designer Collaboration and AI Agents — a Practical Experience
39mWhat's new in Android Studio's AI Agent?
38mDesign and Implementation of Next‑Gen App Development Tools Using Compose Multiplatform × AI
42mWhat's new in Android Studio's AI Agent
26mFigma to Code: AI Generated UI Code with Your Design System
06mReimagining designer-developer handoff: Introducing Relay
30mAI Powered Frontend Development Building Better UIs Faster