Vercel CEO reveals his Agentic Engineering Workflow

TL;DR Agentic engineering shifts software development to autonomous sandboxes and high-level system design. This approach increases speed but requires rigorous security auditing and human oversight to manage technical debt.

Security and Sandboxing Are Critical for Safety

• Human-written code contains immeasurable latent vulnerabilities that agents can easily exploit.
• Supply chain attacks are common because libraries like libheif have hidden remote code execution risks.
• Moving the development lifecycle to cloud sandboxes eradicates supply chain attack vectors.
• Sandboxes allow for strict firewall policies and network monitoring to contain damage.

The Engineer’s Role Shifts to System Design and Verification

• Engineers must define axioms, trade-offs, and performance constraints for AI agents.
• Understanding system latency and caching tiers helps engineers guide AI optimization effectively.
• Verification engineering involves setting guardrails rather than writing every line of code.
• Agents handle implementation details while humans focus on architecture and correctness.

Velocity Requires Focus on High-Alpha Problems

• Companies must measure iteration velocity, not just speed or token usage.
• Build only core products that provide disproportionate alpha; buy standard infrastructure.
• Avoid building generic services like email gateways if reliable alternatives exist.
• Use production metrics like error-free sessions to validate product success in the real world.

Personal Workflow Emphasizes Minimalism and CLI Tools

• The speaker uses FX, a minimal CLI tool with a 6-megabyte binary size.
• FX boots as fast as a standard shell to reduce interaction latency.
• Agents operate within file systems using simple markdown instructions (agents.md).
• Embeddable libraries allow developers to integrate agent capabilities into custom applications.

Next Steps

• Adopt cloud sandboxes for all agentic coding tasks to ensure security.
• Implement automated security benchmarks like DeepSceck to audit codebases.
• Use minimal CLI tools to maintain low-latency interactions with AI agents.
• Join the waitlist for Cloud Room at cloudroom.dev to test open-source agent orchestration.