From 500K to 27K tokens: context engineering for production AI agents

Every SRE has sat through an AI demo that looked great until it hit a real incident. The failure is almost always the same: the model drowns in noisy telemetry, burns tokens, and returns a hallucinated answer.
This workshop fixes that. Bring your laptop. We'll cover Context Engineering: why what you feed an agent matters more than the model, and how scoping telemetry cuts token usage from ~500K to ~27K.
Then we build. Using AURA, an open-source agent harness, you'll wire up an MCP integration, define a custom skill, and construct a runbook that keeps humans in the loop. You'll leave with working code you can adapt in your own pipeline.
Attendees will be able to apply Context Engineering to reduce agent token load and hallucination risk;  connect agents to existing tools via MCP without rebuilding your stack, and design runbooks safe for production use.