Global Coherence separates agent reasoning from shared-state enforcement

Three agents can each make a defensible decision and still exceed a shared budget. A verifier can approve a change using an old version of the facts. The problem is not necessarily weak reasoning: it can be the absence of a component that owns the relevant state.
Global Coherence develops a formal account of these failures and tests selected boundaries in agent workflows. It separates what an agent can know from whether individually valid actions fit together into a valid team result.

Missing facts cannot be recovered by more deliberation

The paper’s observation-aliasing theorem concerns situations that look identical to a policy but require different actions. A guaranteed valid choice exists only if every world compatible with the observation permits at least one common action. If those valid-action sets are disjoint, further reasoning over the unchanged observation cannot identify which world the policy is in.
A controlled revision benchmark illustrates the distinction. The same model solves all 40 tasks when the deciding event is visible. Hiding that event leaves the tested reasoning, role and voting approaches statistically compatible with guessing among three options. Restoring one authoritative fact returns the score to 40 out of 40.
Those finite results do not prove that every tested strategy is empirically equivalent to chance. The paper distinguishes its mathematical statement from the smaller experimental panel. The lesson concerns unavailable information, not a claim that reasoning effort is generally useless.

Agreement between neighbors can miss a team-wide conflict

The second failure can survive full information. Local views may agree in pairs without corresponding to one consistent global state. The paper uses translation loops and shared spending as examples: a locally plausible conversion can fail when composed around a loop, and separately legal expenses can break the total budget.
The proposed framework records which participants share state, how representations translate, whether local views can be reconciled, how actions change state and which past distinctions still affect permitted futures. Its mathematical ingredients organize those checks; the author does not claim that every system must implement the same machinery.
Conventional mechanisms can already perform the necessary work. Version checks, dependency engines, shared ledgers and escrow quotas are examples discussed in the paper. When an existing solver owns the complete relevant state, the reported studies find that it matches the more general harness.

Commit enforcement is different from showing a counter

In the reported five-run TeamBench comparison, ordinary teams exceed a shared 20-call budget in every run. Showing the live count still leaves four violations. Enforcing the budget when calls are committed leaves none, without a reported fall in mean task progress in that small comparison.
The distinction is operational: information about a limit is not enforcement of that limit. The paper assigns proposal generation to the model and state ownership, admissibility checks and commit authority to the harness.
It also examines stale reads and silently reverted changes. A check against current state alone can miss that a proposal was based on older information. The framework therefore retains the version or facts a proposed change depended on, rather than assuming the latest state tells the whole story.

The evidence has bounded scope

The agent experiments use one model family and modest panels. The author presents scale as a next step, not an established result. Detecting inconsistency also does not guarantee a cost-free repair: in the ontology-alignment study, the network-level check removes the reported collisions, but the chosen repair reduces mean F1.
For builders, the paper supplies concrete questions: who owns a shared constraint, what did each proposal read, and which check spans the complete dependency? Adding another manager agent does not by itself answer them. That manager also acts from a limited view unless the surrounding system gives it authoritative state and enforceable commit rules.