Local Fetch Improvements and Tavus Fixes for Agentic Web — September 4, 2026

Today, I set out to enhance the local fetch capabilities and resolve some lingering issues with the Tavus integration. In about an hour of focused work, I managed to ship a feature that improves how external calls are rendered, along with crucial fixes for the Tavus interface and the radio functionalities. The work feels like a step forward, but not without its share of challenges.

The main feature I shipped today was the untrackedFetch for render-path external calls. This was a follow-up from my previous work on September 3, where I laid the groundwork for better handling of external requests. Previously, these calls were somewhat cumbersome and often led to delays in rendering content. The new untrackedFetch mechanism aims to streamline this process, making it more efficient and responsive within the agentic web context. I’m excited about how this will improve user interaction on the platform, especially for those utilizing the GIGI AI concierge.

However, with new features come bugs, and I encountered a couple of them while working on the Tavus fixes. The first issue was related to the in-app frames, specifically for the AISEO board and Calendly booking functionalities. They were not displaying correctly, which made it difficult for users to interact with those tools seamlessly. After spending a focused half-hour diagnosing the problem, I realized that the frames were being blocked due to some outdated configurations. I updated the settings to unblock these frames, which should enhance the overall user experience significantly.

The next fix was a bit more intricate. I had to ground the Lucas radio-stage tools within the #SPACE interaction configuration. This was crucial because the radio tools were not utilizing the correct persona, which in this case was the OWNED Interviewer persona. Instead, they were defaulting to a tool-less stock ID. This misalignment not only disrupted the intended functionality but also affected the overall consistency of the user experience. After some digging, I was able to implement a fix that ensures Lucas live now correctly employs the OWNED Interviewer persona, which should make the radio experience much more engaging for users.

Reflecting on today’s work, I appreciate the intricacies of building a platform like this solo. While AI tools like Claude Code in VS Code have been invaluable for managing these complexities, there’s no substitute for the hands-on troubleshooting that comes with real-time coding. I’m continuously reminded that while AI can assist with many tasks, the human touch is irreplaceable in diagnosing and fixing nuanced issues. Each bug I tackle not only teaches me more about the system but also strengthens my resolve to build a robust platform that can stand on its own.

As I look toward the future, I can see how these incremental improvements will contribute to the vision of a one-man-show company with a billion-dollar valuation. Each feature and fix is a building block towards achieving that goal. I’m excited about the potential impact of the local fetch improvements and Tavus fixes on user engagement and satisfaction. It’s moments like these that reaffirm my commitment to this journey, even when challenges arise.

In summary, today’s work was a blend of shipping a valuable feature and resolving critical bugs that enhance the agentic web experience. I’m looking forward to seeing how these changes will play out in the wild and will continue to iterate as I build this ambitious vision into reality.