Mergeable by default: Building the context engine to save time and tokens
summary
Claims a production context engine cuts wasted work and tokens by supplying only task-relevant organizational context, resolving conflicts, enforcing permissions, and personalizing retrieval; a measured case drops wall time from 2.5 hours to 25 minutes and token use from 21M to 10M tokens. It wires GitHub/PR data, Slack/Teams, Sentry/DataDog and MCP servers into a distilled knowledge graph, ships a social graph builder, and reports real gains while warning against cached answers and naive RAG.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 1:18What is a context engine
- 8:47Three myths about context engines
- 10:54Naive RAG over documentation
- 12:18Larger context windows solve everything
- 14:01Satisfaction of search in agents
- 16:10Six requirements for context engines
- 23:37Hard lessons building context engines
- 25:45Where context engines add value
- 31:18AI agents as team members
- 32:06Q&A
- 33:30Context engines vs LLMs and wikis
- 35:16Engineering vs business focus
- 36:22Governance and knowledge synthesis
- 38:07How context engines work
- 40:55Human feedback and conflict resolution
- 41:59Workshop setup
- 42:52Social engineering graph demo
- 1:01:56Roadmap for context engines
- 1:18:46Live PR review with Unblocked
- 1:21:36Demonstrating value
- 1:22:39UX for context engines
- 1:26:39Memory storage and retrieval
- 1:28:07Benchmarking success
- 1:30:35Token usage optimization
- 1:33:00Real-time data updates
- 1:36:40Customer privacy and on-prem
- 1:39:53Cloud vs on-prem
related talks
31mBuilding the Context Engine AI Agents Need
15mUnderstanding Context Engineering for Agentic Systems
27mContext is the new code: The context development lifecycle
31mBuild Intelligent AI Agents With Context Engineering
23mAI won’t fix developer productivity (unless you fix context first)- 27mContext is the new code: The context development lifecycle
19mContext-Aware AI Dev : Next Wave of Engineering Productivity
24mBuilding Queryable O11y with the Model Context Protocol