What LangChain's Memory Playbook and Notion's Model Router Teach PMs

LangChain published a concrete playbook for keeping long-running agents useful as context piles up, Notion's engineering team (led by Ghita Houir Alami) revealed exactly how they measure model quality at scale with their new Auto router, and Dare Obasanjo's numbers on Meta Muse's adoption give us the clearest data point yet on whether consumers actually want agents doing things versus just chatting. ### LangChain - Context Engineering for Deep Agents: How to Remember Without Keeping Everything in Context **Source:** https://www.langchain.com/blog/context-engineering-deep-agents-remember **Credibility:** High (first-party technical writeup from the team building the Deep Agents harness, with specific implementation patterns) **What happened:** LangChain published a detailed breakdown of how long-running agents degrade as their context window fills up, and the specific techniques their Deep Agents harness uses to prevent it: offloading, summarization, skills, and subagents. This follows a companion post from the day before on retrieval techniques for the same harness, suggesting LangChain is building out a full "context engineering" doctrine rather than a one-off feature. **Key technical details:** - **Offloading**: moving information out of the active context window into external storage that the agent can query back into context when needed, rather than keeping everything resident - **Summarization**: compressing completed subtasks or conversation history into condensed representations instead of carrying the full transcript forward - **Skills**: packaging reusable procedures/knowledge as callable units rather than re-deriving them from scratch each session (this connects to Anthropic's "Agent Skills" donation to the Agentic AI Foundation from earlier this week) - **Subagents**: delegating bounded subtasks to separate agent instances with their own scoped context, so the parent agent's window doesn't absorb the subagent's working memory - Published alongside a retrieval-focused companion post the previous day, framing this as a two-part system: get the right context in (retrieval), then manage it well over time (memory) **Why it matters for PMs:** If you're shipping any agent that runs longer than a single turn, "it gets dumber the longer it runs" is the failure mode you will hit, and most teams discover it in production rather than in design. This gives you a vocabulary and a set of concrete levers (offload vs. summarize vs. delegate) to diagnose which failure you're actually seeing before you reach for "just use a bigger context window," which is usually the wrong fix and the most expensive one. **Critical questions:** - At what point does summarization lose information the user actually needed later? Is there a way to measure "summarization regret" before a user complains? - How do these four techniques trade off against latency? Offloading and retrieval both add round-trips that a chat-only agent doesn't pay. - Is this specific to LangChain's harness, or are these patterns portable to agents built on other frameworks (Mastra, raw API calls)? **Action you could take today:** If you have an agent in production that handles sessions longer than ~10 turns, pull a transcript where a user complained about degraded quality and map which of these four failure modes (stale context, lost summary, no skill reuse, no subagent delegation) actually caused it.