Agent Context Protocols Enhance Collective Inference

Arjun Beniwal, Devansh Bhardwaj, Shreyas Chaudhari, Ashwin Kalyan, Tanmay Rajpurohit, Karthik R Narasimhan, Ameet Deshpande, Vishvak Murahari

Keywords: Agent Protocols, multi-agent systems

Abstract: Multi-agent systems hold the promise of enabling \emph{autonomous AI assistants} capable of solving real-world problems that span many interdependent steps. Yet existing protocols, such as agent-to-agent (A2A) messaging, stop at standardizing communication---they do not provide the sustained coordination required for \textbf{long-horizon, complex tasks}. We introduce \textbf{Agent Communication Protocols (ACPs)}, a unified framework that elevates multi-agent interaction from message passing to \emph{reliable execution}. At their core are \emph{execution blueprints} (persistent DAGs that capture task dependencies and intermediate states) and \emph{fault-tolerant recovery mechanisms} (standardized status codes and context-rich error handling). Together, these allow agents to coordinate, replan, recover, and sustain progress to complete complex long-horizon tasks. Empirically, ACPs set a new state of the art on \textbf{AssistantBench} (28.3\% accuracy), generate \textbf{human-preferred multimodal reports} (outperforming baselines in $\approx$85\% of evaluation dimensions), and demonstrate robustness in ablation studies on a synthetic dataset curated for benchmarking long-horizon, complex tasks. Beyond these results, ACPs mark an initial step toward enabling AI assistants to coordinate multiple agents on complex tasks over long horizons.

Archival Option: The authors of this submission do *not* want it to appear in the archival proceedings.

Submission Number: 101