Complexity and interpretability rarely coincide: complex systems are usually too opaque to question, while transparent ones are too simple interesting dynamics to emerge.
What if we treated LLM agents as the basic computational unit to build agentic substrates? Paper thread 🧵
No Digg Deeper questions have been answered for this story yet.
Because the agents interact in natural language, the collective remains open to inspection: you can read the traces, or simply ask the agents themselves! Where classical ALife derives interpretability from unit simplicity, agentic collectives derive it from unit interrogability.
Because the agents interact in natural language, the collective remains open to inspection: you can read the traces, or simply ask the agents themselves! Where classical ALife derives interpretability from unit simplicity, agentic collectives derive it from unit interrogability.
What do we mean by *agentic* anyway? We consider LLMs to be agentic when endowed with persistent memory, a self-extensible pool of tools and skills, and the capacity to act unprompted. Easy.
In short: we argue that collectives of agentic LLMs can be seen as a computational substrate with individual LLMs as the basic computational unit. The resulting substrate displays emergent dynamics absent from isolated models, and, we argue, can serve as an experimental testbed.
Link to pre-print: https://arxiv.org/abs/2607.01047
Joint work with Ane Espeseth, @nisioti_eleni, @risi1979 and @stenichele
