From Coordination to Authority

MAST organized observable breakdowns into fourteen failure modes across system design, inter-agent misalignment, and task verification.^[1] Six experts developed the taxonomy from 150 traces; most of the roughly 1,600-trace corpus was annotated by an automated pipeline. Agreement on applying the refined taxonomy establishes consistent use, not that the categories are exhaustive, mutually exclusive, or uniquely causal.

Holding weights fixed while varying scaffolding moves scores by roughly 10–20 points on a fixed coding subset.^[2] Harness-Bench reports 5,194 trajectories across eight model backends and six configurable harnesses while preserving traces, artifacts, validator outputs, token counts, permission violations, and process-quality signals.^[3]

The execution condition is a model–harness–tool–environment configuration, including routing, permissions, budgets, timeouts, and fallback behavior. Complete-harness comparisons reveal sensitivity but do not isolate the mechanism behind each effect.

Silo-Bench isolates a communication–reasoning gap: agents exchange messages successfully while failing to reason over information distributed across participants.^[4] Coordination needs a usable representation of current ownership, dependencies, evidence, blockers, commitments, corrections, and completion criteria. Raw conversation may contain those facts without making them current, salient, or binding.

TraceElephant’s best static configurations reach roughly 65.9% attribution accuracy at agent level but only 30.3% at step level. Removing task inputs degrades step-level accuracy by up to 76% relative to full information.^[5] MP-Bench shows that several causal perspectives can remain defensible for one trajectory,^[6] while VerifyMAS tests whether a hypothesized error is present, explanatory, and attributable to a component or step.^[7]

MAS-FIRE introduces known disturbances and measures propagation, detection, silent failure, defensive response, and recovery, with recovery varying by topology.^[9] Controlled intervention strengthens evidence about the tested mechanism without establishing production incidence.

VERIMAP moves checks into decomposition so intermediate outputs can be blocked before downstream work inherits them.^[10] A deterministic oracle, hidden test, learned evaluator, same-context critic, human review, and formal proof still establish different things.

Overlaying Governance models principals, recursive delegation, contextual scope, attenuation, expiry, and revocation as a compositional authorization layer.^[11] Formal permission correctness shows that a system enforced a policy. It does not show that the decision was sound, adequately supervised, or legitimate.

The Six Sigma Agent executes atomic, independently verifiable actions repeatedly across diverse LLMs, clusters the outputs, and takes the winning cluster. It reports 3.4 defects per million opportunities at thirteen agents. Its headline 14,700× improvement is measured against single GPT-4o-mini execution at 50,000 DPMO, the weakest baseline evaluated; the reported 80% cost reduction is relative to more expensive models, with roughly 47% added latency at five samples.^[12]

By contrast, From Spark to Fire models collaboration as a dependency graph in which minor inaccuracies are reused until they harden into false consensus. A single injected error seed produces widespread failure; its governance layer prevents final infection in at least 89% of runs.^[8]

The papers describe different systems. Six Sigma assumes statistically independent errors, per-action error below one half, and failures that do not converge on the same wrong answer. From Spark to Fire studies agents that consume one another’s outputs, where shared context is the collaboration mechanism.

Independence is necessary but not sufficient. It makes errors distinguishable; voting or verification removes them. N-version programming found decades ago that independently written programs fail on correlated inputs more often than independence predicts.^[18] Shared specifications created common-mode failures there; LLM pipelines add a direct inheritance channel through generated context.

Across 500 sequential experiments, ten domains, three heterogeneous models, and 1,250 responses, Hallucination Cascade reports normalized hallucination falling from 0.422 at the first agent to 0.272 at the final agent. Factual accuracy also falls, from 0.789 to 0.769.^[16]

Our interpretation is that the cascade may be converging on claims that are harder to mark wrong rather than adding information. The accuracy decline is too small to establish that mechanism by itself; the divergence between metrics is the evidence.

The studies estimate different statistics. From Spark to Fire asks whether an injected error reaches the final artifact, a tail question. Hallucination Cascade measures average inconsistency. Smoothing and hedging may improve the average without protecting against one inherited false premise.

Across 260 configurations, performance relative to a single-agent baseline ranges from +80.8% on decomposable financial reasoning to −70.0% on sequential planning. Trace-level errors amplify 17.2× under independent parallel execution, 7.8× under decentralized coordination, and 4.4× under centralized coordination.^[13] Here “independent” means parallel outputs aggregated without voting or verification. Independence makes errors filterable, but only aggregation or verification filters them.

Across seven benchmarks, a single agent role-playing a homogeneous workflow matches multi-agent performance while reusing its KV cache. OneFlow cuts GPT-4o mini inference cost from $0.198 to $0.026 on HumanEval and from $2.343 to $0.819 on MATH.^[14]

A weaker vendor-authored study reports silent degradation increasing with chain length.^[17] Its proprietary evidence and exponential “entropy principle” do not support a general law; the observation only reinforces the need to test depth explicitly.

Coordination helps when work is decomposable and verification is centralized. It hurts when work is sequential, context is shared, and no independent check sits between stages. Whether errors can correlate is the moderator, yet almost every paper in this corpus leaves it implicit.

Benchmarks that assign one responsible agent and step assume a privileged cause exists. MP-Bench shows that several attributions can remain defensible for one trajectory.^[6] If failures are multi-causal, single-label ground truth rewards models for resolving ambiguity confidently rather than correctly.

An output can be checked by a verifier that shares the producer’s failure modes. Structural independence may require a separate process or model family, context, evidence path, criterion, and verdict the producer cannot override. Claims of independent validation should report criterion origin, evidence separation, calibration, and blocking power.

A correct, verified output is not thereby permitted to change external state. Reliable action requires a principal, delegation path, bounded scope, authorization state, executor, and revocation or recovery. Authorization research has begun formalizing this,^[11] but benchmarks rarely record who could act or under what scope.

A transcript is not an institutional record. Reconstruction may require the approved artifact version, criterion and verdict, authority state, executed action, resulting external state, and later corrections. Provenance research has begun mapping execution evidence,^[15] but retention alone does not let a successor recover what was committed.