Decentralized Coordination Protocol of AI Agents

One day, our cyberspace will be flooded with AI agents. Before that day, we need to figure out how to coordinate our AI agents better. Let’s say each of us own 5 agents, there will be 8 billions*5, 40 billion agents in the future. How should my AI assistant interact with yours? What if they collude? Many companies already start bringing AI agents to our day-to-day life.

Now is the right time to call for a new decentralized coordination protocol. Before large language model (LLM) based agents become the new paradigm, multi agent systems rely on reinforcement learning to model complex interactions and achieve coordinations. The general architecture of these systems is new, marked by an increasing reliance on the sovereignty and autonomy of individual agents. Therefore, LLM-based multi agent systems require new coordination mechanisms.

We want to improve Language Agents as Optimizable Graphs, a computational graph for collective intelligence (intelligence arises from interactions among individuals). We envision a more generalized and optimal protocol. There are three core ideas/insights.

  1. This target protocol does not connect node optimizations with edge optimizations. Node optimizations are mostly LLM prompts, which should affect orchestration (edge optimizations). What if agents build their own graphs autonomously?

  2. The target protocol mainly focuses on problem-solving, and demonstrates optimization results on related benchmarks. AI agents will not only serve as problem solvers; as demonstrated by Character AI, they also impact companionships and provide emotional support. Since the target protocol is based on swarm, which perform well in world simulations, we aim to develop a more generalized coordination protocol that can address diverse and complex inter-agent relationships.

  3. An agent’s capabilities are upper-bounded by the complexity of the world it lives in.
    Games (and simulation in general) will provide the next trillion high-quality tokens to train our foundation models. In addition, unlike multi agent reinforcement learning systems, which primarily rely on offline training datasets, LLM-based multi agent systems learn mainly from real-time feedback through interactions with the environment, other agents, and humans. So we aim at having our proof of concept in a gaming environment for people to interact with an AI agent society.

What is your discovery methodology for investigating the current state of the target protocol? Eg: field observation, expert interviews, historical data analysis, failure event analysis

In what form will you prototype your improvement idea?

  • Have our agents in a customized version of this open source MMORPG game: it is similar to Minecraft, where agents can take actions and evolve themselves in a complex virtual worlds, as well as team up with people.

Hi Christy,

Thank you for shedding light on such a fascinating area of study!

Your idea reminds me of a compute governance system, particularly one that prompts different levels of entity reactions based on a safety level alert system (like the ASL level system by Anthropic), resonated with me. I’ve been pondering a similar concept, focusing on the crucial role of AI agent interactions within a complex system. This includes micro-level interactions (AI agents with each other and with users) and macro-level interactions (LLMs with each other and with labs).

I envision a dynamic network where agents and users serve as nodes, and the potential for interaction determines the strength of the connections. The system’s global activation level would influence these interactions, streamlining processes such as data queries and information sharing without the constant need for manual permission grants. By leveraging an agent’s coordination system, which assesses authenticity and qualification based on historical permissions and metadata, we could vastly improve efficiency.

The interaction between AI agents can be initiated through various scenarios, such as when one agent needs to access another’s data, cite a paper, or invite an expert to a conference. This suggests the existence of an underlying network where agents and users act as nodes. These nodes are connected by links that represent the potential for interaction, weighted by similarity and the likelihood of interaction, all governed by a system-wide activation level.

The current model, which requires User A to manually grant permission for Agent B to access specific documents each time, is inefficient. To streamline this process, we propose the development of an agent coordination system. This system would allow User A to delegate certain decision-making powers based on predefined criteria, such as the authenticity of Agent B, their qualifications, and a history of permissions granted by User A. The aim is to automate the permission process, relying on a weighted assessment of the connection between nodes to facilitate seamless, authorized data access without requiring manual approval for each instance.

My background as a data scientist and research on the knowledge graph has equipped me with skills in literature review, data ETL, model deployment, and network analysis. I’m keen to explore how my expertise could contribute to developing this system further. I would be thrilled to discuss the possibility of collaborating on this project. Thank you for considering!

1 Like

I sent you a email! Excited to connect!

Thanks for replying! looking forward to chat with you:)

Hello Christy, over in the Discord that I am part of for the EISPP project LLMs are frequently mentioned for development of knowledge graphs because it is expected to be onerous to expect users to create RDF on their own. These graphs describing people and projects are expected to self-aggregate.

I’ve been working on this project for a long time, and the Discord started at the beginning of the year. I put in a late PIG submission.

The Discord link is here: EISPP / Peer Production

earlier transcript from a call: January 17 2024 Kickoff: · bshambaugh/eispp · Discussion #26 · GitHub
( I think in this case we were meeting on Zoom)

Feel free to ask more questions. Maybe there is a point of collaboration? Just curious. I cannot claim to be an AI expert. (taking a deep dive at a hackathon doesn’t count :P)

1 Like