Agents at Work: OS primitives, orchestrators, and project RAG

Microsoft 2.5: EVP Pavan Davuluri wants to remake Windows for both human and agent users reports Microsoft redesigning Windows with native support for agentic workloads — new identity/security primitives and Execution Containers to run human-plus-agent computing. This matters because OS-level execution and identity primitives change how you sandbox, authenticate, and govern agents in production — build your agent architecture to leverage Execution Containers and threat models aligned with Principle 07 (Build the Island).

How to use Cursor Projects details Cursor Projects’ coordinator agent that plans, delegates subagents, and maintains shared context and subscriptions across multi-PR work. This matters because the coordinator-agent pattern is a practical orchestration model for multi-agent delivery pipelines — adopt shared context boundaries and durable subscriptions if you want reproducible, auditable agent workflows (Principle 09).

A deep dive into LangChain and LangGraph walks through LangChain’s model/tool abstractions and LangGraph’s durable TypeScript workflows with branching and human approval. This matters because durable workflow graphs plus explicit human-approval gates turn ephemeral agent runs into auditable, composable pipelines you can test, version, and validate in production (Principles 06 and 15).

Retrieval augmented generation (RAG) for projects announces Claude auto-enabling RAG for projects to expand project knowledge capacity up to 10× while keeping responses fast and accurate. This matters because project-scoped RAG is how you scale grounding and context for outcome-driven agents — plan memory tiers, freshness, and citation surfaces to keep agents truthful and verifiable (Principle 02 and 11).

Spotlight: Startup vet launches Latch to liberate humanity ‘from doing work that owns us’ describes Latch converting narrated task recordings into knowledge graphs so agents understand and automate how work actually gets done. This matters because a knowledge-graph context layer turns tacit operational knowledge into structured, queryable signals for agents — invest in instrumentation that captures task intent and edges if you want reliable automation and transferability (Principle 11).