# [[Live - TCW 06 - Context Engineering Fundamentals]] <iframe width="560" height="315" src="https://www.youtube.com/embed/Vn04wQ0T9PU" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> Related:: "[[Ben Sully]], [[Ioanna Armouti]], [[Tiffany Jernigan]], [[Context]], [[Context Engineering]]" ![[The Context Window 6- context engineering fundamentals.png]] ## Topics Context and context engineering: what counts as context, the context window as an attention budget, Assistant Search and semantic retrieval, embeddings, evaluating retrieved context, the tool tax and tool deferral, progressive disclosure, long-term memory, and why removing 80% of a system prompt can change nothing. ## Timestamps ``` 00:00:00 Introductions 00:02:47 The problem: growing context 00:06:37 What is context? 00:13:46 The context window as an attention budget 00:20:36 Assistant Search and how embeddings work 00:38:58 The tool tax: 55,000 tokens before you start 00:44:23 Select, compress, isolate, and tool deferral 00:51:14 Removing 80% of a system prompt, and what's next ``` ## Resources ### Mentioned in the episode ``` ObservabilityCON 2026, Oct 19-21, Pier 27, San Francisco https://grafana.com/events/observabilitycon/ Anthropic on Claude Code context growth (the 2.6x and 324:1 figures) https://claude.com/blog/claude-opus-5-5-built-for-coding-sessions-that-use-more-context Anthropic: Effective context engineering for AI agents (the attention budget, and the "smallest possible set of high-signal tokens" definition) https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents Anthropic's tool search docs (55,000 tokens before any work, over 85% reduction, and tool selection degrading past 30-50 tools) https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool Anthropic: The new rules of context engineering for Claude 5 generation models (removing 80% of Claude Code's system prompt) https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models LangChain: Context engineering for agents (write, select, compress, isolate) https://www.langchain.com/blog/context-engineering-for-agents The Context Window #4: MCP vs CLIs for AI https://youtube.com/live/ZR6U4TN8-Cw gcx https://github.com/grafana/gcx Grafana MCP server https://github.com/grafana/mcp-grafana word2vec, the paper behind the embeddings example https://arxiv.org/abs/1301.3781 ``` ### Go deeper on what we discussed ``` Ioanna's write-up of Assistant Search, including infrastructure memories https://medium.com/grafana-labs/search-but-deeper-how-grafana-assistant-finds-the-right-context-c3a1e27beea5 Cyril Tovena on progressive tool discovery (the source of the 40,000 token figure, and the facade approach that replaced it) https://medium.com/grafana-labs/progressive-tool-discovery-battling-tool-overload-in-ai-agents-418fb0b4cd76 Ward Bekker on designing gcx for agents https://medium.com/grafana-labs/clis-in-the-age-of-agents-how-we-designed-gcx-cdfca8954b30 Alexander Sniffin on the Assistant Investigations harness, including compaction https://medium.com/grafana-labs/inside-the-harness-how-grafana-assistant-investigates-incidents-9a982b8ff01d Drew Breunig: How long contexts fail (the origin of context poisoning, distraction, confusion, and clash) https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html Context Horizon, Grafana Labs' publication on AI and context https://contexthorizon.substack.com/ ``` ### To fill in ``` Nicole's AI Week catch-up video: [URL needed] ``` ## Transcripts ![[tcw 06 - transcript]]