LLMs have achieved remarkable breakthroughs in reasoning, insights, and tool use, but chaining these abilities into extended processes at the scale of those routinely executed by humans, organizations, and societies has remained out of reach. The models have a persistent error rate that prevents scale-up: for instance, recent experiments in the Towers of Hanoi benchmark domain showed that the process inevitably becomes derailed after at most a few hundred steps. Thus, although LLM research is often still benchmarked on tasks with relatively few dependent logical steps, there is increasing attention on the ability (or inability) of LLMs to perform long range tasks. This paper describes MAKER, the first system that successfully solves a task with over one million LLM steps with zero errors, and, in principle, scales far beyond this level. The approach relies on an extreme decomposition of a task into subtasks, each of which can be tackled by focused microagents. The high level of modularity resulting from the decomposition allows error correction to be applied at each step through an efficient multi-agent voting scheme. This combination of extreme decomposition and error correction makes scaling possible. Thus, the results suggest that instead of relying on continual improvement of current LLMs, massively decomposed agentic processes (MDAPs) may provide a way to efficiently solve problems at the level of organizations and societies.

Most people in the field know that models usually fall apart after a few hundred steps because small errors just keep adding up until the whole process is ruined. The paper proposes a system called MAKER which uses a strategy they call massively decomposed agentic processes. Instead of asking one big model to do everything they break the entire task down into the smallest possible tiny pieces so each microagent only has to worry about one single move.

For their main test they used a twenty disk version of the Towers of Hanoi puzzle which actually requires over a million individual moves to finish. They found that even small models can be super reliable if you set them up correctly. One of the main tricks they used is a voting system where multiple agents solve the same tiny subtask and the system only moves forward once one answer gets a specific number of votes more than the others. This acts like a safety net that catches random mistakes before they can mess up the rest of the chain.

Another interesting part of their approach is red flagging which is basically just throwing away any response that looks suspicious or weird. If a model starts rambling for too long or messes up the formatting they just discard that attempt and try again because those kinds of behaviors usually mean the model is confused and likely to make a logic error. By combining this extreme level of task breakdown with constant voting and quick discarding of bad samples they managed to complete the entire million step process with zero errors.

And it turns out that you do not even need the most expensive or smartest models to do this since relatively small ones performed just as well for these tiny steps. Scaling up AI reliability might be more about how we organize the work rather than just making the models bigger and bigger. They even did some extra tests with difficult math problems like large digit multiplication and found that the same recursive decomposition and voting logic worked there as well.