Multi-Agent Coordination, Tool Grounding, and the Trust Problem

Survey: Trust and Authorization Boundaries in Multi-Agent Systems

As agents delegate tasks to other agents, the question of who trusts whom becomes a real security problem. This paper maps the attack surface and proposes a framework for scoping authority between orchestrators and subagents.

ToolBench-2: Grounding LLM Agents to Real-World API Calls

Tool use is still the biggest gap between demo agents and production agents. This benchmark tests whether agents can reliably select and chain real APIs under noisy, ambiguous documentation conditions.

Hierarchical Planning Reduces Hallucination in Long-Horizon Agent Tasks

Breaking tasks into subtask hierarchies cuts the rate of mid-trajectory hallucinations significantly compared to flat planning. Practical finding for anyone building agents that need to execute more than a few steps reliably.

When Agents Disagree: Conflict Resolution Strategies in Collaborative LLM Systems

Multi-agent systems frequently reach contradictory conclusions, and most frameworks just pick the last output. This paper benchmarks five conflict resolution strategies and shows that structured debate outperforms voting on factual tasks.

Adaptive Tool Retrieval Outperforms Static Toolsets in Agentic Workflows

Giving agents a fixed toolset is a design smell. This paper shows dynamic tool retrieval at inference time improves task success rates by over 20% on complex workflows, with only modest latency cost.