Tavus Enhancements & Agentic Internet Updates — June 6, 2026

I just finished an almost 23 hour straight development session where I made some major improvements to both the Tavus platform and the agentic internet framework. I've been concentrating on perfecting the user experience and performance of the systems since this is imperative for constructing my envisioned one-man company valued at a billion dollars.

For example, one of the biggest changes made was repositioning the visitor self-view picture-in-picture (PiP) Tavus interface window from the bottom right to the top right corner. Although it seems like a small change, it greatly helps usability as it allows users to view themselves while interacting with the platform, rather than having to look away. Changes like these really do add up and I’m discovering that user experience is the most important aspect of building a successful portal CRM.

Also, I have completed Phase 3 updates for the SEVEN and Claude integration on the backend. This includes a long-running worker that manages the Claude Messages API agent loop, which now has support for Read, Write, Edit, Bash, and Finish tools. I've added a safety model and auto-deployment feature that allows for changes to be made without impacting the user experience. This was very time consuming as I had to ensure that the HARD_STOPS and TSC gate were going to work as intended to prevent any undesired results during deployment.

The biggest challenge came from advances in the agentic internet framework. I developed a per-portal MCP server that supports spatial queries and introduces a new discovery integration. Also noteworthy is the SEVEN-to-Claude work queue which enables requests to code change more effciently. This task was not just time-consuming, it was more demanding than I had expected. I had to deal with a number of issues, ensuring the approval gate and run dispatcher worked as intended. I hit a few testing dead ends, but my debugging helped clarify the flow and give me a better implementation.

My focus was also on improving the embed parity features. I mounted the PortalHostSiteFrame and improved the Tavus agent integration so it can control the website, calendar, and form tools. The updates I made to the SEVEN system prompt were also important and reflect the Tavus tools, memory, and data layer updates. These changes are essential for the GIGI AI concierge to function without issues.

One more critical change I made was setting the visitor camera to ON by default. I believe this will enhance visitor engagement. Combined with a persistent per-visitor memory feature, this will create a more tailored experience. I discovered that advocating for a sales-oriented system prompt can influence user actions a lot, so I designed a playbook aimed at booking calls or lead captures in a matter of minutes. This is crucial for an effective AI website builder, and is in line with my future aspirations.

I had a lot of successes today, but I ran into some challenges as well. One of the challenges I faced was a bug that took a ton of time to figure out. It had to do with the interaction configuration race condition that I encountered. I had to do an extensive debugging process, which included tracking changes to interaction configurations and assuring that my logging captured every message sent by the app. It was a very good learning opportunity and reminded me to do more thorough testing in the future.

Looking back at the end of the day, I feel tired, but also I feel proud of what I accomplished. I get closer to my vision for the agentic web with each feature that is implemented. I am proud to say that I have built most of the platform entirely on my own with Claude, my AI copilot. I enjoy the challenge, and every line of code I write improves my company’s worth to $1 billion. Even though the work is tough, the result is what keeps me motivated.