Fable vs 10 LLMs: Who Better Handles the God Node from a Real Agent

The author from Habr in July 2026 conducted a detailed experiment: took a god node from a real working LangGraph agent and organized a two-stage tournament for 11 language models — five American and six Chinese, including Fable from Anthropic. Additionally, they applied three independent analysis methods to understand which participants can be trusted.

Why god node is a fair test

In LangGraph projects, a god node is a monolithic node that takes on too many responsibilities: parsing input data, making decisions about tool invocations, processing results, forming final output. Such nodes grow gradually: at the start of a project, responsibility boundaries are not yet clear, and later it becomes scary to change the architecture — and the node continues to expand.

Refactoring a god node is a task where an LLM cannot guess the "correct" answer based on statistical patterns. Here you need real understanding of architecture:

• Understand the graph as a whole and find hidden dependencies between nodes
• Propose a specific decomposition scheme with justification for the chosen boundaries
• Not violate the existing contract: inputs, outputs, side effects
• Take into account the peculiarities of LangGraph — state transfer, conditional edges, breakpoints

Beyond architectural understanding, care is important: refactoring a god node in production code risks touching unexpected dependencies. A model that proposes "just split into several functions" doesn't understand the context — and its plan is useless in practice.

The author used not a synthetic example, but real working code with real history. This is crucial: real code contains randomness, historical artifacts, and non-obvious dependencies that don't exist in textbook examples.

How the two-stage tournament worked

In the first phase, each of the 11 models received a god node and an assignment: propose a specific decomposition plan. Among the American participants is Fable, the latest Anthropic model at the time of publication. Six Chinese models represented the largest AI laboratories in China.

In the second phase, each model received proposals from the other ten and had to evaluate them: find strengths and weaknesses of each solution, choose the best one, and justify the choice. This makes it possible to measure two skills at once — the ability to generate architectural solutions and the ability to critically analyze others.

Cross-audit is a methodologically non-standard technique. In typical benchmarks, a model receives a task and produces an answer, but it's unknown whether it understands what a good solution is in this area. When a model evaluates competitors, it becomes clear how much its understanding of the "correct" answer aligns with real engineering standards.

After collecting all proposals and mutual assessments, the author applied three independent analysis methods. The results partially coincided, partially diverged — and this is exactly what became the main conclusion of the experiment. From here comes "Death of the Gods" in the title: flagship models with the best reputation did not provide consistently superior results. According to the author, synthetic benchmarks and real engineering code are two different tasks, and leaders in one don't always turn out to be leaders in the other.

What this means

Position on public leaderboards poorly predicts how a model will handle code review or architectural proposals in a specific project. The gap between "the best model by rating" and "the best model for your stack" can be substantial.

Practical conclusion: before delegating technical decisions to a specific LLM — especially in refactoring or architectural review tasks — it's worth testing it on tasks from real code. The two-stage format with cross-evaluations from this experiment is a good template for such testing: it reveals not only generative capabilities, but also critical understanding of the subject area.