X / Twitter
Box CEO Aaron Levie
The AI agent stack is evolving at a startling pace. An architecture developed over half a year just 6 months ago may already be obsolete; starting over can be better than painstaking repairs. Levie identified a recurring pattern: each technical leap makes an old capability layer redundant while creating new capabilities needing scaffolding. With RAG pipelines, for example, much carefully tuned retrieval can now be replaced by agentic search as context windows expand. At the same time, stronger tool use enables more complex tasks, creating new demand for sandboxed code execution. This cycle of one layer disappearing and another emerging is likely to continue for years. Levie’s conclusion: teams without sufficient speed and agility will be on the back foot.
https://x.com/levie/status/2035171720945115469
Y Combinator CEO Garry Tan
Garry Tan shared AI workflow techniques he uses. He no longer prompts with “continue” or “keep going,” hinting at a more efficient way to advance LLM generation. He also introduced GStack’s /plan-design-review tool: iteratively improve a mediocre design or wireframe rated 4 out of 10 until it reaches 10. The core method asks the LLM to score it first, then specify “what is missing to make it a 10.” This structured prompt works well for design review. He also joined community discussions about Claude Code, reflecting continued attention to AI coding tools.
https://x.com/garrytan/status/2035197298586857711
https://x.com/garrytan/status/2035214231625474489
https://x.com/garrytan/status/2035227471273238749
Thariq of Claude Code @Anthropic
Thariq recommended Anthropic’s agent SDK as the easiest current way to build agents, sharing relevant links. He is also trying to build a development livestreaming system so others can watch progress in real time, though he admitted he has not worked out how to broadcast publicly without being “instantly owned.” Together, the posts show his recent focus on building and sharing Claude Code-related agent infrastructure.
https://x.com/trq212/status/2035372729726087456
https://x.com/trq212/status/2035491380923916506
FirstMark Capital VC Matt Turck
Matt Turck reflected on whether moats can be built in the AI era. Citing LinkedIn (2003), YouTube (2005), and Twitter/X (2006), he noted how network effects create extraordinary longevity. His question is whether AI products can build similar flywheels, perhaps through data loops. That remains unclear, making the observation relevant to investors and founders assessing long-term AI competition.
https://x.com/mattturck/status/2035458987940880466
Builder Zara Zhang
Zara Zhang observed a counterintuitive phenomenon: almost every heavy AI user she knows has become busier and more anxious rather than more relaxed. People expected 10 times the productivity to let them finish earlier; instead, that 10-fold productivity created 20 times the to-do list because so much more is possible. The post resonated widely, receiving more than 1500 likes and 150 reposts, touching a common but underacknowledged paradox: the stronger the tools, the harder boundaries become to maintain.
https://x.com/zarazhangrui/status/2035251996564103267
Every CEO Dan Shipper
Dan Shipper announced a $5000 purchase of Figma stock, citing optimism about SaaS companies adapting to AI and Figma’s currently low share price. He specifically named Figma CEO Dylan Field (@zoink), saying he would not miss the AI opportunity. This is a personal investment decision and reflects some builders’ optimism about traditional SaaS adapting to AI.