Correcting Noise-Mispecified Operator Selection in Wild Compound LLM Agents

Compound LLM agents deployed in the wild compose reasoning, coding-style, refinement, stopping, and tool-like operators. Their reliability depends not only on whether each operator works in isolation, but also on whether the controller can correctly assign credit when one operator is flaky. We identify a controller-level reliability failure mode. Under heterogeneous operator noise, including decoding variance, parsing brittleness, sandbox or executor failure, and backend drift, aggregate reward-gradient updates propagate the failure signal of one unreliable operator into every co-selected operator, biasing future routing decisions even when each operator is individually fine. We formalize this setting as \emph{Partial-Observable Operator Selection (POOS)} and show that score-function controllers propagate the full coalition-noise budget into every coordinate update. The correction is not a larger search space but a different learning statistic: paired marginal differences cancel cross-operator noise, and pooled marginals attenuate it by $1/(4K)$. POISE instantiates this principle with REINFORCE exploration, pooled marginal attribution, and singleton estimates fused by inverse standard error. Across real replicate-noise evaluations, controlled stress tests, reliability stress tests for wild agent operators (Programmer timeout, EarlyStop misfire, dual flaky modules), and ablations, POISE improves in the predicted asymmetric, non-singleton regimes (e.g., $+8.7\%$ real-noise mean improvement, $+27\%$ on the heteroskedastic reasoning subset, and $+10$--$17\%$ on flaky-operator failure channels) and ties or loses when self-noise, homogeneous noise, or singleton optima dominate. The result is a regime-level reliability correction for wild compound LLM agents, not a universal robustness claim.