Sandboxes in Production
summary
FLORA runs Vercel Sandbox in production to let AI write and run code safely: media Actions execute FFmpeg/ImageMagick stacks inside snapshots that spin up in a few hundred milliseconds, network-locked and returned only after strict output validation; the MCP server shrinks 18 one-shot tools to a two-tool model that lets agents write TypeScript against the Flora SDK and run it in sandboxes; Remotion-backed timeline exports use stateful sandboxes to resume renders and cut latency while preserving deterministic assets.
AI-written summary · about
chapters (8)
- 0:00FLORA: a creative operating system
- 1:55Buttons are rigid, AI is squishy
- 4:29When to let the agent cook
- 5:44Workflow 1: Actions and the Fauna agent
- 8:14Workflow 2: the FLORA MCP server, 18 tools to 2
- 10:57Workflow 3: a timeline editor on Remotion and sandboxes
- 13:22The pattern: harness, validate, ship
- 14:06Turning the arbitrary into the reliable
related talks
24mContained Intelligence: Secure Sandboxing for Gen AI Code
25mPlease Sandbox your AI Agents
12mDocker Sandboxes for AI
29mSandbox for Agentic Application With Cloud Native Stack
28mSandbox: Run Cline In the Cloud
26mrm -rf: Horror Stories From Unsandboxed AI Agents (and How Docker Fixes This)
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
40mArrakis: How To Build An AI Sandbox From Scratch