newsfilter
← Interviews
AI Engineer22 min

What it takes to set intelligence free — Fabian Hedin, Lovable

Fabian Hedin

Conference Presentation
  • Strategic Pivot on Agentic Systems: Lovable's co-founder Fabian admits the company almost failed in 2023 by over-indexing on complex agentic systems before LLM intelligence was sufficient, initially shipping a simple one-call-per-message model before realizing the need for sophistication.

    • Current Status: As of late 2024/early 2025, the company has returned to a more sophisticated agentic architecture, now running production systems processing over 1 billion tokens per minute.
    • Core Thesis: The platform prioritizes user outcomes (business creation, monetization, value) over code management, positioning itself as a tool where users do not need to care about the underlying code structure.
  • Platform Capabilities and Architecture: Lovable provides a unified interface where a chat describes intent and a preview shows the output, supporting complex software creation without traditional coding workflows.

    • Agent Runtime: The system utilizes a separate inbox for each agent, allowing interaction from humans, other agents, or scheduled system triggers.
      • Trajectory History: Agent history is stored as an append-only log inspired by Git, allowing for forking (parallel agents) without duplicating shared history, though merging is not supported.
      • Stateless Processes: The runtime separates the agent's task history from the executing process, enabling zero-downtime deployments (700 times per day) while long-running agents continue uninterrupted.
      • Prompt Flexibility: By decoupling the raw event history from the prompt, the system allows for dynamic summarization strategies optimized for different LLM context lengths and capabilities.
    • The "Computer" (Sandbox Environment): The platform uses a checkpointing system to spin up development environments in approximately 230 milliseconds.
      • Golden Checkpoint: Each project starts from a pre-loaded "golden checkpoint" containing the OS, system services, and a dev server.
      • Resource Optimization: Implementing copy-on-write for RAM and streaming memory allows for near-zero memory overhead when forking environments (e.g., for "Drafts" features) and boots machines before full memory allocation is complete.
      • Performance Tuning: Lovable forked the Vite dev server and rewrote it in Rust, reducing startup time from 14 seconds to 3 seconds and decreasing memory footprint by 6.5x compared to the Node.js version.
  • Verification and Feedback Loops: Long-running agents (up to 10+ hours) rely on rigorous feedback loops to prevent divergence and ensure quality.

    • Inner Loop (Per-Turn): Verification occurs during execution via build checks, lint/type error detection, runtime error scanning, and headless browser exploration of the generated app.
    • Outer Loop (Between Turns): The system integrates business metrics and user feedback to optimize the product post-deployment.
      • Business Integration: The platform controls payments, email, SEO, and advertising, allowing the agent to optimize based on conversion rates and transaction data.
      • A/B Testing: The system supports automatic A/B testing of copy and design changes to measure impact on user behavior.
  • Scalability and Future Outlook: The architecture supports agents working without time limits, facilitated by the separation of task history and compute processes.

    • Production Scale: The system currently supports multiple agents sharing a single computer instance to explore different solutions simultaneously.
    • Forward-Looking: Lovable intends to continue expanding capabilities to orchestrate LLMs into complex, reliable agentic systems at scale, moving away from the limitations imposed on agent runtimes in the past.