What actually counts as “context” for an AI agent — and how does it find the right things to put in a limited window? In this episode of The Context Window — Grafana’s livestream series on AI in observability — I dig into context engineering fundamentals with Ben Sully and Ioanna Armouti. We treat the context window as an attention budget, walk through Assistant Search and semantic retrieval, embeddings, and how you actually evaluate whether the retrieved context was any good.

What we cover

• What counts as context — and why the context window behaves more like an attention budget than just a size limit.
• Finding the right context — Assistant Search, semantic retrieval, and embeddings: how an agent surfaces what it needs instead of being handed everything.
• Evaluating retrieved context — because retrieval that looks relevant isn’t the same as retrieval that is.
• The tool tax and tool deferral — every tool definition costs context, so when do you load it versus defer it?
• Progressive disclosure and long-term memory — giving the agent what it needs, when it needs it.
• Why removing 80% of a system prompt can change nothing — a nice illustration that more context is not automatically better context.

Resources

• My notes on Context Engineering
• My notes on Context
• My notes on Grafana Assistant