Paper Argues Agentic AI Is Bounded by Environment Determinism

A short position paper out on arXiv makes a claim worth chewing on if you are building anything that chains agent calls together. The argument from Liang Ding and Xintong Wang is that long-chain agent execution fails exponentially in environments designed for human tolerance: with per-step determinism δ < 1, k-step chain success degrades as δ^k. Put plainly, even small amounts of non-determinism in the tools an agent calls compound brutally across a long workflow.

The frame they propose is that environment determinism is a binding axis on agentic AI progress, sitting alongside the more familiar frictions in the AGI-to-ASI scaling debate: the data wall, the abstraction barrier, the embodied bottleneck, and multi-agent trust. The class of tasks they want to cover is the one where outcomes are verifiable economically, physically, or through multi-party settlement, which is a fairly specific scope and worth keeping in mind when reading the paper's stronger claims.

What they actually deliver are three formal results: a Determinism-Efficiency Bound on chain-task success, a Verifier-Goodharting Floor on flywheel ceilings under imperfect rewards, and a convergence condition for environment-side skill evolution. On top of that they propose an operational layer, a Supply Certainty Index over five measurable properties and a five-level Determinism Maturity Model presented as an adoption ladder. They also lay out an open-question programme with explicit null results that would force retraction, which is a more honest move than most position papers make.

The honest caveat is that the abstract is what is publicly visible, and the δ^k story is a clean theoretical bound, not an empirical measurement. The reporting doesn't give you per-tool δ numbers for any real platform, nor how retry logic or self-verification interact with the decay, so the practical implications stay open. The paper itself engages with three competing positions: sim-to-real sufficiency, alignment sufficiency, and AI-as-normal-technology.

For practitioners, the useful takeaway is the inversion. If the bound holds even loosely, the reliability of an agent system is mostly determined by the determinism of the environment you ask it to act in, not the cleverness of the model on top. That points at a different procurement and design discipline than waiting for the next model, and at tool and platform vendors who can credibly sell determinism as a feature.