Should compression be aggressive or conservative in different environments?

ACON's guideline optimization runs as two steps — utility maximization (UT) and compression maximization (CO) — and the paper's discussion draws a practical rule from how each behaves across its three benchmarks: "Applying only the utility maximization step (UT) improves performance while reducing token cost across all benchmarks, whereas the compression maximization step (CO) further lowers token cost but may slightly hurt accuracy depending on the environment." The two steps are not equally safe everywhere, and the paper ties the choice to what kind of environment the agent is working in rather than to any property of the agent itself.

The source gives the rule in its own terms: "in verbose and noisy environments such as AppWorld, where observations contain redundant API outputs and distractors, UTCO is often preferable because aggressive pruning can improve both efficiency and focus. In contrast, in high-fidelity information-seeking tasks such as OfficeBench and 8-objective QA, UT is generally safer because over-compression may remove subtle facts needed for final decisions." The paper states this plainly as a recommendation — "UTCO is recommended when observations contain substantial redundancy, while UT is preferable when precise fact retention is critical" — framing aggressiveness as a property to be matched to the task's information structure rather than applied uniformly.

This is a second axis on top of Can an external manager handle context for frozen agents?, which also lands on a trade-off between aggressive and conservative compression but locates the deciding variable in the agent: stronger agents want high-fidelity preservation, weaker ones need aggressive compression to stay reliable. ACON's excerpt does not frame its UT/CO choice in terms of agent strength at all — the same guideline-optimization framework is run per-benchmark, and the resulting recommendation tracks redundancy and fidelity requirements in the environment's observations, not the capability of whichever model is acting on them. Read together, the two sources suggest compression aggressiveness may need to account for both the agent's reliability and the environment's noise level, though neither paper's excerpt tests the two factors jointly. The relation to Can prompt optimization cut agent token usage without retraining? is that both describe the same guideline-optimization method; this note isolates the environment-dependent part of its results.

What the excerpt does not establish is how to choose UT versus UTCO without already knowing, from benchmark results, which regime a new environment falls into — the recommendation reads as a post-hoc characterization of AppWorld versus OfficeBench and Multi-objective QA, not a diagnostic a practitioner could run on an unfamiliar environment in advance. It also leaves open whether the same per-environment dial could be set automatically, or whether it still requires a human or a separate classifier to decide how noisy or fidelity-sensitive a new agentic environment is before applying one step or both.