After studying hundreds of failure patterns in long-running agentic workflows, the argument is shifting from giving LLMs more responsibility to giving operational state more authority.Hey, my smart friend. It was a little quiet here over the past few weeks. The blogs temporarily disappeared,…
Annons
Annons
After studying hundreds of failure patterns in long-running agentic workflows, the argument is shifting from giving LLMs more responsibility to giving operational state more authority.Hey, my smart friend. It was a little quiet here over the past few weeks. The blogs temporarily disappeared, together with my usual commentary on whatever Big Tech decided was cool to demo that day. But there was a reason: I was working on a bit of research.Yeah, I know, it’s a terrible content strategy, obviously. Like drawing six glowing boxes and then connecting them with arrows and writing “agentic architecture” underneath would have been considerably faster.The work was all about the question if language models really the right computational foundation for long-running enterprise autonomy.My view is moving steadily towards “ehh, naaa probably not.”LLMs are extraordinary at processing language, no doubt, but the trouble starts when we take that capability and promote it into something for the enterprise to run workflows. We then expect a probabilistic model, with an ambiguous language-based world model, to maintain operational reality, but the process keeps changing underneath it.➡️ Read this blog: I’m completely done with LLMs in the enterprise | LinkedInMOPPING WITH THE TAP RUNNINGThe first result is a meta-study, “Mopping with the Tap Running”, into the failure modes of LLM-based enterprise workflows.What we found is that the number of issues in the ATLAS corpus related to long-horizon agentic workflows had increased nearly tenfold in the last year, and that was all about multiple seemingly small errors accumulating and causing workflows to break.➡️ You can download the MWTR whitepaper here. Read Mopping with the Tap Running.➡️ And if you’re interested in the data or access to the corpus of 4100 papers, drop me a DMWe also presented the findings and the argument for moving away from language as the operating world for complex workflows.➡️Watch the presentation of the results on YOUTUBECOORDINATION FMThe research then led to a more fundamental question if language itself is the wrong substrate for coordination. That question led to a new foundational model for agentic workflows called COORDINATION FM (CFM, preprint).The underlying idea is that an enterprise system should know where it is at any given moment (State). But many current agentic architectures reconstruct their situation whenever the model is invoked. They rely on language to rebuild an operational world that should have existed independently of any conversation in the first place.The CFM is based on state.What is the current state, and what change is actually allowed from here?And when an operational state exists explicitly, then the system doesn’t need to rediscover reality every time that it reasons because it can determine if an action is valid before it carries it out.➡️ Read the CFM preprintCFM BOOKWe also wrote a book on how to implement this in practice. The supporting training material and reference architecture are available.WHAT COMES NEXTThe next step in the pipeline is a novel architecture we call ATOMIC, short for Autonomous Task-Oriented Machine Intelligence & Control. The architecture explores the proposition that is at the centre of our work: that AI does not have to be probabilistic.We will publish the ATOMIC architecture shortly in the form of 11 interconnected papers, and peer reviewers are very welcome, particularly those who are able to find out where the idea breaks. It is a collection of 11 architecture papers, built together with the folks at Evolver.AI — the first deterministic Enterprise Brain to hit the ground running.This is a short overview of the concepts behind ATOMIC:➡️ Download the whitepaper➡️ Read this blog about: The day the enterprise grew a brain | LinkedInSo yes, after spending an unreasonable amount of time studying how these systems behave once they are allowed to wander around an enterprise, I increasingly suspect that progress will come from giving language models less responsibility instead of continuously finding new ways to give them more, and patching the results with more band-aids.LLMs are extraordinary at language, but that does not automatically make language the right place to store operational reality and govern authority or decide if the next action is actually permitted or not.The direction we are exploring is therefore quite different.Language will remain the interface through which humans and machines communicate. State will become the representation of what is actually happening, and perhaps Control Science determines which transition may occur next and if the evidence is sufficient to allow it, which, considering how enthusiastically we handed a text generator the keys to the kingdom, feels less and less like a revolution. The true revolution lies in deterministic state-based intelligence, and Evolver.ai and hopefully ATOMIC/CFM will be the front-runners.And now we get to find out whether the bloody thing works or not.Signing off,MarcoFor my daytime job, I’m an agentic AI factory builder and a researcher at NCC-1701, a research lab operating at the frontier of unsexy AI. In the evenings, I write about why building this stuff is a lot harder than the papers would have you believe.This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!State-based intelligence is nigh was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.