Lovable co-founder and CTO Fabian Hedin describes the runtime and sandbox behind its coding agents in this 22-minute CODE Summit talk on AI Engineer.
Complexity failed before it worked
- GPT Engineer began with one model call per user message in 2023.
- A specialist multi-agent framework made the product worse; a simpler system restored responsiveness and reliability in late 2024.
- Hedin says stronger models now support longer runs, backed by durable execution and feedback.
Keep task history separate from prompts and workers
- Each agent has an inbox accepting messages from humans, other agents and scheduled systems.
- An append-only event history records actions and tool outcomes, with parent links inspired by Git.
- Agents can fork from a shared history; merging their histories has no defined equivalent in this design.
- Render a model-specific prompt from that history, including context compaction, while retaining the original events.
- Persist between turns so a replacement worker can continue the same task across deployments.
- Hedin reports roughly 700 deployments per day and describes tasks lasting ten hours or more.
Checkpoint and fork the computer too
- A golden checkpoint includes the operating system, tools, development server, memory and running processes.
- Hedin reports about 230 milliseconds to clone that baseline and subsecond resumes of project checkpoints.
- Save checkpoints after turns, suspend idle sandboxes and fork environments for parallel experiments.
- Copy-on-write memory shares pages until a process changes them; streamed restoration starts useful work before all memory arrives.
Shrink the preview engine
- Lovable’s OJ implements Vite-compatible frontend tooling in Rust, reducing startup work and memory pressure.
- Hedin separately cites twice-faster cold starts, previews falling from 40 seconds to three, and roughly 6.5× lower memory use.
- The company’s earlier OJ report measures different endpoints: median total load of 17.4→8.0 seconds and sandbox acquisition of 14.5→3.0 seconds.
- OJ’s compatibility paths can still use Node sidecars for plugins or server rendering, as its repository documents.
Close both feedback loops
- Within a turn, check builds, lint, types, runtime errors and browser behavior.
- Between turns, use product feedback, visitor behavior, experiments and payment-conversion data to guide the next task.
- Hosting, logs, authentication and business tools give the agent places to act and observe results.
Caveat: timing, deployment and token-volume figures are Hedin’s vendor claims. The talk supplies no common benchmark tying its different speed figures together.
“In the end it didn’t work at all. We just made the product worse.” — Fabian Hedin, 2:56