Tavus Enhancements & Agentic Internet Updates — June 6, 2026

Today was a marathon of development, clocking in at around 23 hours, but I shipped some significant updates to both the Tavus platform and the underlying agentic internet framework. I focused on enhancing the user experience and optimizing system performance, which is crucial for my vision of building a one-man-show company with a billion-dollar valuation.

The most notable changes I made include moving the visitor self-view picture-in-picture (PiP) from the bottom right to the top right of the Tavus interface. This seemingly minor adjustment actually enhances usability, ensuring that users can easily see themselves while engaging with the platform. Small changes like this add up, and I’m learning that user experience is everything in building an effective portal CRM.

In the backend, I finalized the Phase 3 updates for the SEVEN and Claude integration. This involved a long-running worker that manages the Claude Messages API agent loop, which now supports Read, Write, Edit, Bash, and Finish tools. I implemented a safety model and an auto-deployment system that allows changes to be made seamlessly without disrupting the user experience. This took a considerable amount of time to get right, as I had to ensure that the HARD_STOPS and TSC gate were functioning properly to prevent any unintended consequences during deployment.

One of the bigger challenges was the updates to the agentic internet framework. I created a per-portal MCP server that allows for spatial queries and integrates a new discovery mechanism. The SEVEN-to-Claude work queue was also a significant addition, allowing for more efficient processing of code change requests. This task was more complex than I initially anticipated, especially as I had to ensure that the approval gate and run dispatcher were functioning correctly. I encountered a few dead ends while testing, but debugging helped me clarify the flow and ultimately led to a more robust implementation.

I also focused on enhancing the embed parity features. I mounted the PortalHostSiteFrame and improved the integration of the Tavus agent, ensuring it has full access to website, calendar, and form tools. The recent updates to the SEVEN system prompt were also essential, as they now reflect the latest capabilities of the Tavus tools, memory, and data layers. These improvements are vital for the GIGI AI concierge to operate smoothly and effectively.

Another significant update was making the visitor camera ON by default, which I believe will enhance engagement. Coupled with a persistent per-visitor memory feature, this allows for a more personalized experience. I learned that pushing for a sales-focused system prompt can greatly influence user behavior, so I created a playbook that encourages booking calls or capturing leads within a short timeframe. This approach is essential for building an effective AI website builder, and it aligns perfectly with my long-term goals.

While the day was filled with various successes, it was not without its challenges. One bug that ate up a lot of time was related to the interaction configuration race condition I encountered. It required a comprehensive debug process where I had to trace interaction configuration changes and ensure that my logging captured every app message. This was a valuable learning experience, teaching me the importance of rigorous testing before deployment.

As I look back on today's work, I feel a mix of exhaustion and satisfaction. Each feature I implemented brings me one step closer to achieving my vision for the agentic web. I’m proud of how far I’ve come building this platform almost entirely solo, with Claude as my AI partner. I’m committed to this path, knowing that every line of code brings me closer to my billion-dollar valuation goal. The grind is relentless, but the progress is evident, and that’s what keeps me going.