A breakthrough in parallelization depends on efficiently orchestrating long-running loops and short-running tasks, using the new tools discussed in part 2. What is the main challenge as we push these limits? The biggest issue with long-running loops today is context collapse, which invariably occurs when something goes off track in the agentic workflow. Even when everything works as intended, LLMs still struggle, with a decline in the LLM's attention and intent detection often attributed to context rot.

So much can change in one year, and even more in just six months. The season of agents is now in full swing, and no one is building for chat anymore. In this post, we'll discuss parallelization - what current capabilities exist and what may come next. None of the predictions mentioned here are inevitable, and each will require significant effort to materialize, some may never pan out. However, if they all stack together, it can spell the beginning of a whole new season of agents.

Having worked on bringing some form or another of C++ compilers into customers hands during my years at Microsoft, news of Anthropic writing a C compiler from scratch with AI attracted my attention for all the reasons you would expect. But what stood out to me has nothing to do with that C compiler and has all to do with what’s coming next in the season of agents.