Solving a Million-Step LLM Task with Zero Errors (2511.09030v1)

Abstract: 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.

• The paper proposes a Massively Decomposed Agentic Processes (MDAP) framework that breaks tasks into minimal microsteps with systematic error correction.
• The MAKER system achieves error-free execution on million-step tasks by leveraging redundant voting and red-flagging to mitigate error propagation.
• Empirical results demonstrate that even less advanced LLMs perform reliably at scale, with analytical models guiding cost and reliability trade-offs.

Massively Decomposed Agentic Processes: Error-Free Million-Step Reasoning with LLMs

Introduction and Problem Setting

The paper presents Massively Decomposed Agentic Processes (MDAPs) as a novel paradigm for achieving error-free, large-scale reasoning with LLMs. It confronts the persistent inability of LLMs to maintain reliability over long, sequential tasks, formalized in the catastrophic failure of monolithic agents on benchmarks such as Towers of Hanoi with increasing sequence length. The central thesis is that scaling up error-free LLM execution cannot be achieved by mere improvements in base model "intelligence," but instead requires architectural modularity, fine-grained task decomposition, and systematic error correction.

MDAP Framework and MAKER Implementation

The MDAP framework relies on three synergistic components:

1. Maximal Task Decomposition (MAD): Tasks are decomposed into the minimum unit steps, each assigned to a specialized microagent.
2. Step-Level Error Correction: Output for each step is subject to voting procedures (first-to-ahead-by-$k$), leveraging independent samples from the base LLM to promote semantic consistency and reduce stochastic error propagation.
3. Red-Flagging for Reliability: Outputs evidencing pathology—e.g., excessive length or format violations—are recognized as signals of confusion and systematically discarded, reducing both step error rates and cross-step error correlation.

These techniques are embodied in MAKER, a system capable of executing over a million sequential agentic LLM steps (Tower of Hanoi with 20 disks), achieving zero total errors.

Theoretical Scaling Laws and Error Correction

Rigorous analysis quantifies the scaling properties of MAD under different error correction and decomposition strategies. Key equations show that the solve probability and API cost scale favorably only under extreme decomposition, with the number of required voting rounds per step growing logarithmically with the total number of steps, and the total cost growing as $\Theta(s \ln s)$:

Figure 1: MAKER error-free solve rate scaling laws as a function of per-step error rate and voting margin $k$; high reliability is achievable with modest cost even as sequence length grows.

Further analysis establishes sharp cost and reliability contrasts between MAD and coarser decompositions:

Figure 2: Required voting rounds $k$ and total API cost scale logarithmically and log-linearly, respectively, with task size when maximal decomposition is employed, supporting scalability.

Figure 3: Expected cost increases super-linearly with agent step granularity—assigning more steps to each agent leads to impractical cost escalation, reinforcing the necessity for MAD.

Empirical Benchmarking and Model Selection

MAKER's performance is validated empirically across a broad array of LLMs (proprietary and open-source). Notably, the per-step error rates of smaller, non-reasoning models are comparable to more advanced reasoning models for tightly scoped steps, refuting assumptions about the necessity of state-of-the-art models for reliable execution:

Figure 4: Comparison of model error rates and projected cost; per-step error rates remain stable as solution length increases, and non-reasoning models deliver more cost-effective error-free runs for million-step tasks.

Model selection for cost minimization is formalized by estimating per-step success rate $p$ and sample cost $c$, enabling scaling millions of steps at practical cost.

Large-Scale Execution and Convergence

With model/hyperparameter selection guided by empirical error and cost estimation, MAKER achieves the first perfect solution of the Tower of Hanoi with over one million steps. System dynamics exhibit sharp exponential decay in the number of undecided steps:

Figure 5: Visualization snapshot of MAKER executing the million-step Tower of Hanoi—illustrates process progress.

Figure 6: Exponential convergence of undecided steps to zero across voting and sampling rounds; bulk of computation is concentrated in initial rounds, as predicted by scaling laws.

Impact of Red-Flagging and Analysis of Failure Modes

Red-flagging drastically reduces both individual error rates and the probability of correlated errors occurring across steps. Empirical paper shows abrupt increase in error rates once response length exceeds a threshold, but overall solve cost remains unaffected due to the rarity of these events. More critical is the reduction in "collisions" (multiple incorrect outputs for the same step), which could otherwise undermine voting reliability in long-horizon runs:

Figure 7: Red-flagging mitigates pathological error sources—excessive output length and format violations—both reducing error rate and decorrelating errors across votes.

Vote races for pathological steps are explicitly visualized, demonstrating robust eventual correctness:

Figure 8: Vote race for a pathological step requiring ≥18 votes; illustrates resilience of statistical error correction.

Broader Implications and Future Directions

Agentic Modularity and Microservices Parallels

The microagent architecture supports modularity, independent scaling, and real-time robustness reminiscent of microservices architectures. This enables parallelization, targeted agent specialization, and system flexibility across roles and tasks.

Limits and Applicability of Decomposition

The universality of MAD depends on the problem's decomposability into small steps solvable by current LLMs. Some domains may lack sufficiently granular, context-isolated steps, limiting the approach's direct applicability.

Error Correlation and Deeper Decorrelation Strategies

Although red-flagging and parallel sampling suffice to decorrelate errors for Towers of Hanoi, more complex cases may require explicit decorrelation (prompt paraphrasing, agent diversification, context randomization). The importance of error decorrelation intensifies with system scale and heterogeneity.

Safety, Control, and Model Welfare

Extreme decomposition supports robust sand-boxing, reduces collusive risks, and enables systematic auditing, offering a safer route to reliable superintelligent deployments with minimal requirement for specialized, high-risk LLMs.

Extending MDAPs to Insight Generation

Current experiments focus on execution; future work will address decomposition strategies where insight generation (e.g., automated step discovery, non-trivial subproblem definition) is needed. Preliminary evidence suggests recursive agentic voting can generalize to more open-ended tasks (e.g., multi-digit arithmetic).

Conclusion

The paper provides a rigorous pathway for deploying LLM-based agentic systems at unprecedented scale and reliability, substantiating that breaking tasks into microsteps and applying efficient error correction enables perfect execution over a million steps, even with inexpensive base models. This advances the field beyond the limits of monolithic agent architectures, opening the door to scaling AI in production workflows, organizational automation, and critical domains with strict reliability requirements. Future research should explore more general forms of agentic decomposition, deeper error decorrelation strategies, and applications in open-ended reasoning, while carefully considering architectural safety and compliance constraints.