Interesting take from @tarunchitra @ 22:04 however, agents being early shouldn't be the point of discussion here Reliability still ranges from subpar to nonexistent in the vast majority of scenarios, especially in scientific research or any type of research involving financial data and so on... Benchmark scores for most “frontier” models keep going up but the reliability issues talked about above persist GPT has web search consisting of web scraping and that doesn’t magically prevent it from outputting degraded results and slop. A browser MCP will not fix that To expand on MCP: it’s a set of tools that unlocks certain things and we are finalizing our own clients, however simply plugging them into LLMs cannot mask their inherent cognition shortcomings. Meaning that it’s nothing more than a tool that needs to be used properly for max performance extraction. So were agents too early or are they doomed to be very limited in the current paradigm? Such opinions have been ours for much longer than the existence of “Chasing AI’s Holy Grail”( )