Abstract:

As large language model (LLM) equips agents with stronger semantic understanding and task execution capabilities, multi-agent systems are transitioning from closed intranets to open networks. They show great potential for cross-institutional financial auditing, distributed scientific computing, decentralized intelligent services, and related scenarios. However, open environments often lack mutual trust, involve heterogeneous interests, and operate without a central authority. As a result, existing systems struggle to support verifiable collaboration, reproducible and recoverable failure handling, and accountable agent behavior. The objective of this study was to systematically review the key challenges of multi-agent collaboration in open environments. These challenges include the tension between the semantic correctness of state transitions and the verifiability of logs, trusted storage and version consistency of long-term memory, and contribution attribution and incentive compatibility among self-interested agents. Based on this analysis, a theoretical framework grounded in blockchain and verifiable computation is proposed. This framework includes an integrated on-/off-chain architecture, verifiable fault-tolerant state machines, verifiable storage and consistency maintenance for semantic memory, manipulation-resistant attribution, and incentive mechanisms. This research will foster deep cross-fertilization among distributed systems, formal verification, game theory, and LLM uncertainty, providing fundamental theoretical support for trustworthy collaboration in the “AI+” initiative.

Key words: multi-agent systems, blockchain, self-healing resilience