Durable State, Failure Taxonomy, and Autonomous Tool Discovery

Tools

LangGraph's Durable Agent State Management Enters Beta

LangGraph's new session management gives agents durable state across conversations. This is table stakes for any production agent, moving us from stateless hacks to persistent, context-aware systems.

Reliable Code Generation for Agent Tooling via Search Methods

The focus is on making agent code generation more reliable. Moving from 'maybe it works' to 'you can count on it' is the difference between a demo and a product.

Research

Evaluating Agents Against Dynamic, Complex Specifications

The research tackles a core builder problem: getting agents to follow complex, long-run specs. If you can reliably test an agent against a dynamic specification, you can actually trust it to do real work.

Agents That Autonomously Discover and Learn New Tools

Instead of hardcoding tools, this research has agents autonomously discovering and learning to use tools. This creates more adaptive, leaner systems that can grow their capabilities.

Analysis

A Taxonomy of Failure Modes in Multi-Agent Systems

This work breaks down why multi-agent systems fail, moving beyond 'they just don't work' to specific failure modes. For builders, it’s a taxonomy of what to debug and how to design for resilience.

News

New Benchmarks Push Agents on Long-Horizon Planning Tasks

Agent benchmarks are maturing beyond simple QA. Evaluating long-horizon planning and decomposition is critical for agents that do complex, multi-step work in the real world.