Tim Berglund and the lightboard are back with an overview of Context Engineering in agentic AI. He breaks down the components that make up context (including system prompts, tools, and resources) and provides a framework for managing these limited resources to build more efficient agents, helping you prepare for the long-horizon future.

You can try Confluent Intelligence at https://www.confluent.io/product/confluent-intelligence 
Promo code: CONTEXTENG

LEARN MORE
► Confluent Intelligence: https://www.confluent.io/product/confluent-intelligence
► Confluent Developer: https://developer.confluent.io

CONNECT
Subscribe, if you dare: https://www.youtube.com/@ConfluentDeveloper?sub_confirmation=1
Community Slack: https://confluentcommunity.slack.com
X: https://x.com/confluentinc
Linkedin: https://www.linkedin.com/company/confluent
GitHub: https://github.com/confluentinc
Site: https://developer.confluent.io

ABOUT CONFLUENT DEVELOPER
Confluent Developer provides comprehensive resources for developers looking to learn about Apache Kafka®, Apache Flink®, Confluent Cloud, Confluent Platform, and any other technology related to the broader Data Streaming Platform. Content on Confluent Developer includes courses, getting started guides, topical deep-dives, patterns, tutorials, and listings of community events. Learn more at https://developer.confluent.io. 

#contextengineering #promptengineering #confluent

YouTube

Prompt engineering is being superseded by context engineering as the critical skill for building agentic AI applications. LLMs are stateless token processors with finite context windows, and as agents iterate over long horizons, six components compete for that space: user messages, system prompts, tool descriptions, resources, assistant messages, and tool call histories. A framework for managing this context includes: writing balanced system prompts (not too vague, not too prescriptive), crafting precise tool descriptions with full schemas, retrieving data intelligently via RAG or MCP rather than dumping everything in, and handling long-running agents through compaction (LLM-based summarization), key-value memory stores, and decomposing complex agents into specialized sub-agents. Filling 60-70% of the context window tends to yield better results than maxing it out.

Prompt Engineering is dead.