The agent race is quietly becoming a retrieval race

This week the interesting news was not a new frontier model. LangChain open-sourced a document extraction service, Smartsheet showed how it built a remote MCP server on AWS, Sebastian Raschka published a deep dive on controlling reasoning effort in LLMs, and the most-read piece on Towards Data Science argued that most companies using AI are blocked by their data platform, not their model. Different vendors and authors, one direction: the bottleneck everyone is attacking is getting the right context to the model, not the model itself.

I see the same thing in enterprise projects. When an agent fails in production, it is rarely because the model could not reason. It is because the agent could not find the right document, the right record, or the right prior decision. The context was in the company all along, just not reachable.

My advice to teams planning agent work for the next two quarters: spend less of the budget on model evaluation and more on making your internal knowledge findable and permission aware. The model you choose matters less than what you can feed it. That is a data engineering problem, and it is solvable today.