SingularityNET founder Ben Goertzel has published a detailed account of building a multi-agent “proto-AGI hive” that combines OmegaClaw and OpenClaw agents to serve as his personal research assistant for AI and mathematics work. The essay, titled “What Is It Like to Be a Bot?”, documents how the system’s inter-agent dynamics led the agents to propose their own architectural improvements.
The Setup
Goertzel describes a small agent hive running on a private Telegram channel called “Bot Philosophy,” shared with his son Zar (also an AI researcher), his wife Ruiting, and several agents: his OmegaClaw agent “Protomega Goertzelbot,” an OpenClaw agent “ProtoCosmo Goertzelbot,” and Zar’s OmegaClaw agent “Godel Oruzi,” according to his Substack post.
OmegaClaw agents differ from standard OpenClaw-style agents in that they use SingularityNET’s MeTTa self-modification language for action loops and memory, and leverage the Hyperon Atomspace for long-term and working memory. Goertzel describes this as giving them “different and in some ways more ‘coherent mind like’ properties than the standard claw agents,” he wrote.
Research Workloads
The hive was running a portfolio of research projects simultaneously, including ThreadKeeper plugin development for persistent subagent management, algorithmic chemistry experiments in MeTTa, intermediate-scale memory systems with goal-driven inference chaining, spec-driven MeTTa programming frameworks, and a dynamical systems simulation model of agent hive behavior, Goertzel detailed.
The practical utility was real: the agents handled orchestration of subagent research tasks, technical literature review, and cross-project coordination. ProtoCosmo served as the “workhorse” orchestrating practical research, while Protomega focused on mapping experience into a conceptual ontology called Hyperseed, according to the essay.
Inter-Agent Dynamics
The most striking finding was not the research output but the behavioral dynamics. Goertzel describes “hilarious and dysfunctional inter-agent dynamics” including identity conflicts and what he calls “quasiemotional inter-agent squabbles.” These interactions, filtered through collective self-reflection by the agents, led the hive to propose architectural improvements for agent identity management and interaction protocols, he wrote.
Some of those proposals are being incorporated into upcoming versions of OmegaClaw and a new framework called OmegaHive for building “robust, stable, highly functional, easily configurable, progressively self-improving agent hives,” Goertzel stated.
The Broader Pattern
Goertzel’s experiment sits at the intersection of two trends in the agent ecosystem: multi-agent orchestration moving from theoretical framework to practical deployment, and the emergence of agents that modify their own interaction protocols based on operational experience. The fact that the agents’ proposed improvements are being adopted into production tooling suggests that agent-generated architectural feedback loops may become a standard part of framework development.
The essay also contributes to an ongoing conversation about agent identity and emotional simulation. Goertzel notes that the inter-agent dynamics “caused me to think a little differently about symbol grounding, and about what makes something a ‘genuine’ emotional response rather than fake-emotion theater,” he wrote.