Post

The State of Context Engineering

The author argues that dependable AI agents need more than larger context windows: they need deliberate context management, increasingly handled by deterministic harness code rather than model prompts.

The talk walks through context poisoning, distraction, confusion, and conflicting instructions, then surveys augmentation, tool selection, quarantine through subagents, pruning, summarization, and offloading. It also describes newer long-running-agent problems such as context inertia. Hacker News readers challenged whether a model’s learned knowledge should be called “context poisoning,” and debated the value of pruning versus cached-context costs; several shared practical habits like manual handoffs and keeping context smaller.