AI Engineer22 min
What it takes to set intelligence free — Fabian Hedin, Lovable
Conference Presentation
- Intelligence is currently deemed sufficient for complex agentic systems, with expectations that capabilities will continue evolving, requiring ongoing readiness.
- The platform philosophy posits that users will not need to manage code in previous ways, aiming to deliver products or businesses rather than just code artifacts.
- Agents are designed to perform arbitrary tasks, including specialized software building, and will operate without length bounds, capable of running for 10 hours or more in an always-on state.
- Hosting primitives including databases, authentication, storage, compute, security scanners, backups, and logs are provided with agent introspective capabilities.
- Operational support extends to payments, email delivery, SEO optimization, advertising on Google and Meta, and product analytics, with agents remaining "in the loop" for these functions.
- The system utilizes an append-only trajectory log similar to Git, allowing for parallel agent spawning at any historical point without duplicating shared history, though merging trajectories is not supported.
- Separating prompts from task history enables the rendering of prompts for any model or task type, while a decoupled process and history design allows for continuous updates to underlying processes without disrupting active agents.
- The platform deploys approximately 700 times daily, roughly once every two minutes, to facilitate frequent updates and maintenance.
- Project initialization from a golden checkpoint occurs in about 230 milliseconds, with checkpoint saving occurring sub-second and machine spin-up happening extremely quickly.
- Resource optimization includes the ability to spin down compute after approximately 10 minutes of user inactivity and spin back up without dropping requests in transit, utilizing copy-and-write on RAM to make forking effectively free in terms of RAM usage.
- Streaming memory capabilities allow machines to boot before all memory is fully ready, while a Rust implementation of Vite reduced code start times from 14 seconds to three seconds and decreased memory footprint by 6.5 times compared to Node.js.
- The system supports running numerous agents in parallel to explore solutions, featuring an inner loop for verification during turns (build success, lint, types, runtime errors, headless browser exploration) and an outer loop for verification between turns using production data and online feedback.
- Automated A-B testing is available to measure changes in conversion, with the system tracking transactions and conversions to optimize business strategy.
- The agentic system currently processes billions of tokens per minute and plans to continue building out its specific capabilities and architecture.