Multi-Agent Coordination, Trust Boundaries, and the Next Wave of Autonomous Planning

AgentRecipe: A Compositional Framework for Building Complex Multi-Agent Systems

This paper introduces a structured approach to composing multi-agent pipelines from reusable building blocks, which is the exact problem most teams hit when scaling beyond single-agent prototypes. If you're building anything with more than two agents talking to each other, this is worth reading.

Trust Hierarchies in Multi-Agent LLM Systems: Formal Threat Modeling

Researchers lay out a formal model for how trust should propagate between agents in a network, including where prompt injection and agent impersonation attacks enter the picture. This is the kind of security groundwork the field has needed before multi-agent goes mainstream in production.

PlanAgent: Long-Horizon Task Planning with Hierarchical Subgoal Decomposition

New work on breaking down long-horizon tasks into verifiable subgoals, with benchmarks showing meaningful gains over flat chain-of-thought approaches. Better planning is still the main bottleneck for agents doing real work autonomously.

Benchmarking Autonomous Web Agents Across Realistic Task Distributions

A new benchmark tests web agents on tasks sampled from actual user behavior rather than curated toy problems, and the performance gap versus previous benchmarks is substantial. Real-world task distributions are messier than researchers assumed, which explains a lot of production failures.