Andrew Qu, Chief of Software at Vercel, explains in an interview with Latent Space that agents represent a fundamentally new form of software compared to traditional web applications due to their dynamic interactions and outputs.
- Qu highlights the development of 'eve', a framework for building agents, which emerged from solving internal challenges faced while building Vercel's v0 product.
- The article outlines key agent primitives identified during this process, including filesystem agents, skills, compaction, and subagents.
- Qu notes that agents are best suited for repetitive tasks requiring reasoning rather than fixed automation, such as legal contract redlining or data queries.
- Vercel aims to provide observability and evaluations out of the box for developers deploying eve, while integrating with specialized partners.
- Skills are emphasized as portable knowledge sources to correct outdated model information, such as steering agents away from deprecated tools like Vercel Postgres.
Qu argues that understanding these distinct requirements is crucial for building effective agent infrastructure that balances autonomy with necessary human oversight.