What a VC Learned Building AI Agents Hands-On

Tom Tunguz from Theory Ventures shares nine practical observations from actually building with AI agents, and the most useful ones cut against the default playbook. He fine-tuned Qwen 3 at 8 billion parameters using reinforcement learning and beat GPT-5.2 zero-shot on well-defined tasks - running locally on a laptop. His multi-agent setup has Claude planning while Gemini and Codex critique the output, turning model rivalry into a debugging workflow. The less obvious insight is about static typing: Rust catches AI-generated code errors that Ruby and Python silently pass through, which meaningfully improves one-shot success rates. Several of these observations - nightly automated prompt optimization, hot-reloading prompt files, consolidating evaluation into closed loops - echo patterns that applied AI teams at startups in Theory's portfolio have converged on independently.