Charles Fleming, Luca Muscariello, Vijoy Pandey, Ramana Kompella
原文摘要
Large Language Models (LLMs) can learn domain-specific languages such as APIs and tool interfaces, enabling agents that act through standardized protocols like MCP. But LLM context windows cannot grow indefinitely, so scaling beyond individual capacity requires agent collaboration: the "Internet of Agents" (IoA). Current network stacks (OSI, TCP/IP) were designed for data delivery, not semantic coordination, so we propose two new layers: an Agent Communication Layer (L8) that standardizes message envelopes, speech-act performatives, and interaction patterns; and an Agent Semantic Layer (L9) that negotiates a shared context, grounds and disambiguates content against it, and provides consensus primitives for population-scale coordination. On two real datasets and two LLMs, L9's schema-grounded disambiguation improves accuracy on ambiguous requests by 13--72 percentage points over blind guessing, at no added cost when disambiguation is not needed; an ablation shows this gain is schema-specific for open-domain QA but not for structured slot-filling, where the protocol's round-trip structure alone already suffices.
Jianwen Chen, Peng Xia, Kaide Zeng, Siwei Han, Zhen-Yu Zhang, Charles Fleming, Xinyu Ye, Huaxiu Yao
原文摘要
Large language model (LLM) agents interact with their environments through executable harnesses that govern tool use and feedback. Optimizing a harness for a fixed policy overlooks how its behavior changes during learning, while training a policy under a fixed harness leaves the interface that generates its training experience unchanged. This motivates adapting the harness throughout policy training. We introduce CoHARL, a population-based reinforcement learning framework that co-evolves a shared policy and a population of executable harnesses. The policy learns from rollouts across the population, and a proposer uses its execution traces to revise harness code during training. CoHARL achieves a 27.6% mean resolved rate on SWE-rebench, improving over harness-only search by 6.4 percentage points. Experiments on SWE-bench Verified and Terminal-Bench 2.0 show gains in software engineering and terminal-based tasks. Under matched training budgets, CoHARL also outperforms searching for a harness once before RL and alternating policy updates with revisions to a single active harness.