The agent framework stack grows up: packaged recipes, hyperedge memory, semantic trace search

Observability becomes the product

LangChain's October 2026 newsletter reads less like a feature list and more like a production checklist. LangSmith AI Search lets teams find conversations based on a semantic understanding of traces, without relying on traditional filters. Engine v2 adds red teaming to catch issues before production, with automated validation of agent fixes. LangSmith Trajectories gives a chronological view of each session across the main agent and its subagents. m:…8529

The newsletter also announced Managed Deep Agents v0.9, with user-level memory, tool use, and task scheduling, and LangSmith Fine-Tuning, which turns production trajectories into training data for fine-tuning. Notably, LangChain published guidance on orchestrating Jev, a decision model, with LangGraph to build agents that are faster and cheaper to run, which connects the framework story to this week's decision-model releases: the frameworks are starting to route around the most expensive model calls. LangSmith Custom Apps lets teams build and publish custom interfaces directly in LangSmith, turning recurring workflows into shared tools on their trace data. LangSmith Trajectories adds a chronological view of each session across the main agent and its subagents. m:…8529

Memory keeps the fact together

Zep shipped v4 with a fix for a subtle memory problem. Most context graphs store each fact as a pairwise edge, a subject and an object, which splits a fact about three or more entities into pairs. The post argues agents can make poor decisions when their context has only part of a fact, so v4 adds hyperedges: a group of pairwise member edges that share one fact, with a minimum of three distinct entities. A hyperedge must contain at least three distinct entities. m:…4088

The mechanics are straightforward: every member edge stores the full multi-entity fact rather than just its own relationship, all members share the same validity window and hyperedge UUID, and a search matching any member finds the full fact. When the fact changes or becomes invalid, Zep writes the change to every member edge. Developers can create hyperedges with the graph.hyperedge.add SDK method, or let Zep extract them automatically from episodes like messages and documents. The extraction setting is on by default for new projects and is one-way, applying only to episodes received after it is enabled. A project-level Enable Hyperedges setting controls extraction, and the setting is one-way. m:…4088

One repo for agent recipes

Google's developer team published a tour of google/adk-recipes, a repository that unifies Agent Development Kit resources into three structured layers, core, plugins, and contrib, serving as a single source of truth for building agents. Developers can use reference patterns from the repository and install modular vertical skills directly into their coding assistants. m:…0046

The vertical skills include retail product search and virtual try-on, and the repository structure is designed to take agents from scaffolding to deployment on Google Cloud infrastructure. In a companion post, Google's developer team argued for replacing subjective vibe checks with continuous evaluation pipelines for multi-turn agents, built on the Gemini Agent Platform, Google ADK, and OpenInference standards. This suggests the frameworks are converging on the same answer: agents need memory that keeps facts whole, tracing that answers plain-language questions, and recipes that make production patterns repeatable. m:…0046 m:…8791