Agent Memory Gets Real, Handoff Patterns Fragment, and The Eval Gap Widens

Tools

LangGraph Ships Managed Short-Term Memory for Persistent Agent State

This is the first serious attempt at managed state that survives across agent runs without builders rolling their own Redis or Postgres layer. For production agents that need to remember context across sessions, this removes a huge chunk of plumbing work.

Analysis

New Research: Multi-Agent Handoff Reliability Drops Sharply After Three Turns

The paper quantifies what builders already suspect: handoff accuracy between agents degrades fast after 3-4 delegation hops. Good empirical backing for keeping agent graphs shallow.

The Batch Highlights Growing Gap Between Agent Benchmarks and Production Metrics

DeepLearning.AI points out that SWE-bench and similar evals still don't predict real-world agent reliability. Builders keep shipping agents that ace benchmarks but fail on messy production tasks.

News

Anthropic Updates Claude Model Docs With Agent-Specific Capabilities Matrix

New breakdown of which Claude models handle tool use, long context, and structured output best for agentic workflows. Useful for builders doing model routing decisions in multi-model setups.