Global coherence: every agent can be right and the team still wrong
Oct 1 Heng: observation-aliasing bound + shared-budget TeamBench—showing the live count still overspends 4/5; commit ownership or escrow hits 0/5. Models propose; the harness owns distinctions and commit.
Source: Global Coherence: When Every Agent Is Right and the Team Is Still Wrong
Xin Heng (Tote AI) names a failure class continuous factories hit constantly: local validity does not imply global coherence. Agents, humans, and mixed teams act on partial views. Two modes dominate. Observation aliasing—worlds that need different actions look identical in context—cannot be fixed by more reasoning, roles, votes, or samples inside the same window (Observation-Aliasing Impossibility Theorem). On a 40-task revision probe, a frontier model scores 40/40 when the deciding event is visible; with it hidden, tested arms land 12–17/40, statistically compatible with chance among three options (1/3), until one sentence restores the fact (40/40 again). Pieces that do not glue—pairwise checks pass while the loop fails—need topology, sheaf/gluing checks, groupoid translation loops, and commit-time versioning, not a smarter manager agent (still bound by the same theorem).
The operational punchline is ownership of shared invariants. On a preregistered TeamBench budget panel (shared 20 shell calls): ordinary teams overspend in 5/5 runs; exposing the live count still leaves 4/5; enforcing the budget at commit leaves 0/5, and fixed per-role escrow quotas also reach 0/5. Progress does not collapse in the small enforced comparison. In τ²-bench Telecom, call-time policy checks tie the coherence harness where a change breaks a stated rule, but when a committed change is silently undone the stock agent scores 0.07 on the benchmark grader while the harness scores 1.00. Where a conventional tool already owns the state (dependency engine, coupled solver), it matches the harness exactly—as predicted.
Why it matters for code review
Overnight FOMO is not “are agents idle?” — it is “does anyone own the budget, the version set, and the loop that would catch a globally impossible merge?” Continuous harnessed throughput compounds when the harness holds distinctions and refuses illegal commits; peak-session mastery that proposes from a partial window cannot invent missing state. Review multi-agent PRs for Heng-shaped contracts: put deciding facts into context, give every cross-agent constraint an owner (ledger or escrow), check state changes at commit with recorded read sets, and do not confuse a manager agent with a harness. Pair with Foremerge’s plan-scope merge tax when worktrees isolate files but not intents, and with Guo’s groundability when weak auditors need independent checks before the judge.