Global Coherence in Multi-Agent Collaboration: Local Intelligence Cannot Substitute for Missing Global State

The paper reframes multi-agent failure as the global coherence problem, a failure of shared state rather than of model intelligence, and gives the Observation-Aliasing Impossibility Theorem that characterizes its exact boundary: a policy can guarantee a valid action exactly when all worlds producing the same observation share an admissible action. Relative to the common view that attributes collaboration failure to reasoning ability or insufficient communication, the theorem shows that if k indistinguishable worlds require pairwise-disjoint actions, the best randomized worst-case success is 1/k, and more reasoning, roles, messages, or samples cannot recover the missing distinction. The boundary is given as a theorem and tested on a controlled revision benchmark: the same frontier model scores 40/40 when the deciding event is visible, tested arms score 12-17/40 when it is hidden, consistent with chance (1/3), and restoring one authoritative fact returns 40/40.

The paper proposes local-to-global runtime semantics X = (H, C, G, F; D): topology H records overlapping scopes, category C governs state-changing actions, groupoid G retains reversible translations, sheaf F tests whether local views glue into one world, and minimal history D keeps only distinctions that alter legal futures; models propose while the harness owns shared state and governs commit. The framework moves ownership of shared state and the commit right from the model to the harness, offering executable runtime semantics for multi-agent collaboration rather than conceptual coordination advice. Nine studies test both the failure and its boundary: on TeamBench ordinary teams exceed a shared budget in 5/5 runs, a visible live count leaves 4/5 violations, and commit enforcement leaves 0/5; on tau2-bench Telecom current-state checks score 0.07 after silent reverts while the harness scores 1.00.

The paper reports a counterintuitive conclusion: local intelligence cannot substitute for missing global state; where a conventional solver already owns the complete relevant state, it ties the harness as predicted. This result locates the harness benefit in the missing-global-state setting rather than claiming general superiority over conventional solvers, providing a boundary condition for when the framework is needed. The conclusion comes from the controlled comparisons across the nine studies, including the controlled revision benchmark, TeamBench, and tau2-bench Telecom, together with the tie when a conventional solver already owns the complete state.

The work targets multi-agent runtimes with overlapping scopes that need shared state, in settings where models propose and the harness owns shared state and governs commit. It enables teams to stay globally coherent when the deciding event is hidden or silent reverts occur, benefiting researchers and engineers building multi-agent collaboration systems. When a conventional solver already owns the complete relevant state, the harness adds no further advantage, and this condition delimits the scope of the results.

A careful reader would still watch how often the 1/k boundary of the Observation-Aliasing Impossibility Theorem arises across observation structures, how harness commit enforcement behaves over longer horizons and larger teams, and how general the tie condition is when a conventional solver already owns the complete relevant state. This reading is based on the abstract only, without figures or full experimental detail, so implementation and statistical specifics remain open questions for the original text.