An InfoQ presentation from Baruch Sadogursky and Patrick Debois argues that many coding-agent failures come from poor context management, not from a lack of raw context-window size.
Their core claim is that “the right 300 tokens” can beat a much larger prompt filled with noisy or stale material. They discuss practical context engineering techniques including lazy-loaded skills, versioned context artifacts, externalized memory banks, and LLM-as-a-judge evaluations.
The point is useful for teams adopting agents in software work. Large context windows make it tempting to dump documentation, logs, policies, and code into every request. But if the model cannot tell which details are authoritative, current, and relevant, the extra text can make answers less reliable rather than more.
The presentation is not a product release, but it captures a workflow shift that is becoming important in AI engineering. Treating context as architecture means deciding what an agent should know, when it should load that knowledge, and how teams verify that the context actually improves the work.