Amjad Masad, co-founder and CEO of Replit, shared insights on AI agent development at SaaStr AI 2026, emphasizing breakthroughs in context window size and agent longevity. Masad revealed that the context window for AI agents has expanded from 16,000 tokens two years ago to over one million tokens today, allowing agents to run practically indefinitely without rebooting, according to saastr.com.
During the event, Masad discussed how Replit’s AI agents, such as 10K (Marketing AI) and QBee (Customer Success AI), operate continuously with effective context compaction. He highlighted the advantage of a mono repository architecture, where multiple applications run under one codebase and URL, enabling agents to retain knowledge from previous builds. This approach mirrors architectures used by Google and Facebook, enhancing efficiency and scalability.
These advancements underscore a shift in AI agent capabilities, enabling more complex and persistent interactions than previously possible. The increase in context window size surpasses human memory capacity, marking a significant milestone in AI development. The mono repo strategy also supports rapid iteration and integration of multiple AI tools, reflecting broader trends in AI infrastructure design.
Replit’s progress was demonstrated through saastr.ai, which runs about 10 applications including a startup valuation tool used over one million times and a pitch deck grader accessed 4,500 times, illustrating practical adoption of these AI agents.