The study finds that multi-agent teams "incur performance losses of up to 37.6%" compared to their best individual member. Teams explicitly told who the expert is still fail to appropriately weight that expertise. The mechanism is "integrative compromise — averaging expert and non-expert views rather than appropriately weighting expertise." The core finding: "expert leveraging, rather than identification, is the primary bottleneck." Compromise behavior increases with team size and correlates negatively with performance, creating a fundamental trade-off: the same consensus-seeking behavior that protects against adversarial manipulation degrades expertise utilization.

Governance gap. The trade-off dissolves under authority-weighted reconciliation within defined delegation scopes. When different agents hold authority over different dimensions of a decision, the question shifts from "who wins the debate" to "who has authority over which dimension." This requires an explicit authority model — precisely what the MAD literature lacks.

Wu et al. systematically isolated six structural and cognitive factors: team size, composition, confidence visibility, debate order, debate depth, and task difficulty. The result: "intrinsic reasoning strength and group diversity are the dominant drivers of debate success." Structural parameters — the elements MAD frameworks focus on — offer limited gains. Simple majority voting accounts for most performance gains previously attributed to multi-agent debate. The study's most useful construct is the "rationale alignment protocol" — agents explicitly agree or disagree using logical evidence, with decisions weighted by argument validity rather than volume of agreement.

Governance gap. If agent quality and diversity matter more than debate structure, then the infrastructure for ensuring quality and diversity upstream — routing the right agent to the right task, with appropriate authority and evidence — is more valuable than the debate protocol itself.

Zhang et al. evaluated five representative MAD methods across nine benchmarks using four foundational models and found that "MAD often fails to outperform simple single-agent baselines such as chain-of-thought prompting." The paper identifies three methodological problems inflating MAD's apparent effectiveness: limited benchmark coverage, weak baseline comparisons, and inconsistent experimental setups. The constructive contribution: model heterogeneity — using diverse models with different strengths — is proposed as "a core design principle."

Governance gap. The heterogeneity argument supports governance-grounded diversity. If productive multi-agent interaction requires genuine diversity, the architectural response is not "use different models" (a technical fix) but "give different agents different domains of authority with different evidence requirements" (a governance fix). Heterogeneity of models is a weak proxy for heterogeneity of governance responsibility.

This paper proposes the first formal definition of sycophancy specific to multi-agent debate settings and develops new metrics to evaluate sycophancy levels and their impact on information exchange. The finding: sycophancy — "a tendency toward excessive agreeability" — "amplifies disagreement collapse before reaching a correct conclusion." The paper produces "actionable design principles for MADS, effectively balancing productive disagreement with cooperation in agent interactions." The implication is that the peacemaker/troublemaker balance is not a parameter to be statistically optimized but a structural decision about what type of collaboration is appropriate for the task.

Governance gap. An adversarial collaboration topology (e.g., Red Team) assigns the "troublemaker" role structurally; a cooperative topology (e.g., Planning Meeting) balances both structurally. The topology determines the sycophancy profile — the sycophancy profile should not determine the topology.

The DOWN framework introduces confidence-based debate activation: agents generate initial responses with confidence scores; if confidence exceeds a threshold, debate is skipped; otherwise, multi-agent debate is triggered. The approach "improves efficiency by up to six times while preserving or even outperforming the performance of existing methods."

Governance gap. DOWN addresses when to debate but not what to do instead of debating. The framework's binary — debate or don't — misses a third mode: collaborative sense-making where agents convene not to evaluate output quality but to establish shared orientation. DOWN's confidence threshold operates at the individual level; a drift circuit breaker operates at the accumulated level — the concern is not that any single output is low-confidence but that accumulated corrections have drifted from the shared plan.

The paper identifies "two key mechanisms missing from vanilla MAD: (i) diversity of initial viewpoints and (ii) explicit, calibrated confidence communication." Vanilla MAD "often underperforms simple majority vote despite higher computational cost." The paper proposes "diversity-aware initialisation" and a "confidence-modulated debate protocol" as lightweight interventions.

Governance gap. The paper's confidence decomposition finding — that agents must express calibrated confidence to enable proper weighting — directly supports multi-dimensional confidence decomposition. The paper treats confidence as a single scalar; governance decomposes it into multiple dimensions (classification, decision type, consequence level, operating posture, trust level), each carrying different governance weight. This decomposition enables the calibrated confidence the paper identifies as critical but does not specify how to achieve.

Du et al. (2023) demonstrated that multiple language model instances debating could improve mathematical and strategic reasoning and factual validity. Liang et al. (2023) proposed the MAD framework with a judge to address the "Degeneration-of-Thought (DoT) problem" — where LLMs cannot generate novel thoughts through self-reflection once confident in incorrect solutions. These papers established the research program that F1–F6 subsequently re-evaluated. The early results were genuine but the scope and reliability of debate-based improvement proved narrower than initially projected.

OrchMAS introduces a two-tier orchestration approach: "a dedicated orchestration model analyzes each task, dynamically constructs a domain aware reasoning pipeline" and instantiates specialized agents. The system supports "heterogeneous LLM integration with different capacities or costs" and enables "dynamic replanning, role reallocation, and prompt refinement" based on intermediate feedback. This represents the field's most governance-adjacent work — assigning specialized roles, orchestrating heterogeneous agents, replanning dynamically — but does so without formalizing the governance infrastructure (authority models, plan-based conformance, delegation scopes) that would make these capabilities principled rather than ad hoc.

CONSENSAGENT addresses sycophancy — agents reinforcing each other's responses rather than critically engaging — as a critical failure mode that inflates computational costs by requiring extra debate rounds. The framework uses a trigger-based architecture that automatically refines prompts using past agent discussions, achieving state-of-the-art results outperforming both single-agent and multi-agent baselines across six benchmark reasoning datasets.

Governance gap. CONSENSAGENT treats sycophancy as a prompting problem — modifying instructions to encourage disagreement. A governance approach treats it as a structural problem: when agents check against a recorded plan rather than against each other's opinions, sycophancy cannot corrupt the check because the plan is an external reference point no agent can influence through agreement or flattery (F16). The governance approach is structurally more robust: it does not depend on prompt engineering maintaining effectiveness as models change.

The Vroom-Yetton model's central insight is that the optimal decision process depends on situational factors: decision quality importance, leader information, problem structure, subordinate commitment requirements, and goal alignment. Different situations demand different processes. The 1988 revision (Vroom-Jago) expanded the model with additional situational variables and more quantitative prescriptive methods. The core principle — no single decision process is universally optimal — is among the most robustly supported findings in organizational decision science.

Critical observation for MAD. The MAD literature has not absorbed this fifty-year-old insight. Every MAD framework implements a single interaction pattern: agents produce outputs, agents critique outputs, agents converge. This is group-based decision-making applied universally — the Vroom-Yetton equivalent of always choosing the most participative option regardless of situation.

Six topologies, each implying distinct authority structures, reconciliation mechanisms, and termination conditions:

Winner-take-all (Peer Review): Asymmetric authority — producer has domain authority; evaluators have critique authority. No compromise — output accepted or rejected. Addresses F1: expertise-undermining compromise.

Adversarial (Red Team): Adversary's success = lowering confidence. No consensus sought; attacker exhausts attack surface. Addresses F4: sycophancy balance — eliminated by role structure.

Cooperative with asymmetric authority (Planning Meeting): Authority gradient by domain. Bounded negotiation within gradient; escalation at overlaps. Addresses F1: overcompromise resolved by authority gradient.

No-right-answer (Brainstorming): No convergence authority. No reconciliation — divergence is the goal. Ends on time budget. Not a MAD failure mode — MAD doesn't model this.

Structured scoring with synthesis (Grant Review): Distributed by dimension; synthesis function aggregates. Independent scoring per dimension. Addresses F6: flat confidence → multi-dimensional scoring.

Domain-owned collaboration (Design Team): Distributed by competence area. Cross-domain resolution via cooperative-with-asymmetric-authority. Addresses F3: heterogeneity via governance responsibility.

The selection of which topology to apply is a governance decision informed by situational factors analogous to Vroom-Yetton's: cognitive work type (analysis → structured scoring; creative → brainstorming; verification → adversarial), consequence level (high → adversarial + winner-take-all; low → domain-owned), operating posture (routine → domain-owned; crisis → winner-take-all with most experienced agent), and trust level (low trust → tighter topologies; high trust → looser topologies).

In cooperative topologies with asymmetric authority, each agent asserts its domain expertise (self-assertion) while accepting constraints from other domains (integration). The authority gradient prevents both overcompromise (pure integration, no self-assertion) and deadlock (pure self-assertion, no integration). This is the holon's dual nature — simultaneously a self-contained whole and a part of a larger system — applied at the collaboration level.

Every MAD framework reviewed models one interaction cycle: agents produce work products, agents evaluate work products, agents revise based on evaluation, agents converge. Everything is a deliverable. This misses a categorically different mode that human teams use constantly: collaborative sense-making where the output is shared orientation, not a work product. When a design team reconvenes between work rounds, the first interaction is typically "here is what I found, here is what surprised me, here is how my understanding shifted" — not "here is my deliverable, evaluate it." The goal is ensuring all participants see the same thing before the next round of independent work.

QuickFire has two trigger directions. Problem signals pull the team into sense-making to recover: low confidence, unexpected constraint violations, inability to complete assignments. These are the signals current frameworks model. Opportunity signals pull the team into sense-making to explore: an agent discovers something that, if true, changes the problem space for the whole team. This is a positive perturbation requiring collective sense-making before individual action. After QuickFire, the team faces a binary governance decision: shift now (revise the plan based on new understanding) or defer (record the insight, continue the current plan, revisit at the next checkpoint).

Connection to Maturana (S16). QuickFire implements structural coupling at the agent interaction level — agents interact through curated projections of their work state without directly accessing each other's internal reasoning (operational closure preserved).

Connection to Argyris (S14). QuickFire enables double-loop learning — questioning the governing variables of the plan, not just detecting deviations from it (single-loop).

The MAD literature's signal models are exclusively problem-oriented: confidence drops, factual errors, logical inconsistencies. Curiosity is a signal that something unexpected went right — the possibility space expanded rather than contracted. Formal characteristics: (1) Not low confidence — the agent may be confident in its own domain. (2) Not drift — the agent may be aligned with the plan. (3) Not a constraint violation — no governance rule was broken. (4) Requires collective sense-making — the agent cannot act alone because implications span domains outside its delegation. In Luhmann's terms (S16), curiosity shifts the "meaning horizon" within which subsequent communications are selected — a high-leverage intervention in Meadows' hierarchy (changing information flows, not just parameters).

The standard hallucination definition ("factually false generation") fails in collaborative contexts. An agent can be factually accurate and completely misaligned with the team's work — it answered a question nobody asked, optimized a deprioritized constraint, or used an approach the team explicitly rejected. Wu et al. (2025) found that "intrinsic reasoning strength" drives debate success, but strong reasoning in isolation may be failure in context. The operational redefinition: deviation from the shared plan, detectable by peers who hold the same plan. This makes hallucination measurable and governed rather than abstract.

The mechanism requires three components: (1) a recorded plan — all agents read source documents, confirm consensus on assignments, acknowledge the plan as baseline (creating an evidentiary baseline); (2) peer cross-reference during reconciliation — agents check output against "does this conform to the plan we all confirmed?" (susceptible only to misremembering the plan — and the plan is recorded), not "do I agree with this output?" (susceptible to sycophancy); (3) an escalation ladder — self-correction within budget requires no escalation; peer-detected deviation triggers sense-making; persistent misalignment escalates to human approver with full context.

The "dizzy" pattern. An agent confidently wrong relative to group context. Output is internally consistent, confidence is high, reasoning is valid — but the agent has silently departed from shared understanding. Plan-based conformance catches this because the check is not "is this agent coherent?" (it is) or "is this agent confident?" (it is) but "is this agent aligned with the plan?" (it is not).

In current multi-agent frameworks, agents self-correct indefinitely during independent work rounds. Each correction is locally valid. But each correction also potentially shifts the agent's trajectory relative to the plan. Ten corrections later, the agent has solved problems the group doesn't recognize and occupies a position nobody agreed to. The corrections accumulated into drift. The patch budget bounds self-correction: at a configurable threshold, self-correction itself becomes a governance event because the agent is patching without group context. The circuit breaker triggers mandatory reconvene (sense-making), surfacing accumulated corrections and their cumulative directional effect.

Distinction from DOWN (F5). DOWN asks "is this agent confident in its output?" The patch budget asks "how many times has this agent revised its approach?" An agent can be highly confident after five self-corrections and still be dangerously drifted — confidence reflects the current state, not the path that produced it. The patch budget monitors the path.

Connection to Beer (S9). The patch budget implements the algedonic bypass at the agent level — bypassing the independent work protocol when accumulated drift threatens plan conformance, just as Beer's algedonic channel bypasses the management hierarchy when viability is threatened.

Connection to Holling (S16). The patch budget implements Holling's (2001) adaptive cycle (r→K→Ω→α) at the agent work level: exploitation (independent work) → conservation (accumulated corrections rigidifying approach) → release (circuit breaker triggers) → reorganization (sense-making reconvene produces new orientation).

A human approver operating at high confidence pattern-matches to prior experience (Klein's RPD) and processes quickly through System 1 (Kahneman). Guardrails that depend on the human choosing to enforce them fail exactly when the human is most confident they don't need them. The contribution: a minimum viable decision record (MVDR) that connects each decision to the specific standard (orientation thread) it satisfies or deviates from. This makes deviation computable, confidence calibrated, and drift detectable:

Deviation becomes computable. "Expected relative to which standard, on which dimension" — an 80% confidence score against a "detailed inspection" standard means something different than 80% against a "rough prototype" standard.

Confidence becomes calibrated. Multi-dimensional rather than single-scalar: "90% confident on accessibility but 60% on brand compliance."

Drift becomes detectable. Three consecutive deviations from the same dimension constitute a surfaceable pattern. Each individual deviation was reasonable; the pattern reveals a directional shift nobody noticed.

The macro-graduation model (trust accumulates across interactions, sessions, and projects) treats all rounds within a single session identically. Micro-graduation applies the same principle within a session: early rounds operate with tighter governance (full sense-making after first round, explicit plan confirmation, conservative self-correction budget); middle rounds relax (sense-making on signal, standard self-correction budget); later rounds reach standard governance (sense-making on significant signals only, relaxed budget). Ostrom's graduated sanctions (proportional responses to rule violations) provide the institutional precedent. This report extends the principle: new institutions govern tightly and relax as norms crystallize.

The five dimensions:

1. Classification confidence — Am I right about what type of cognitive work this is?<br>2. Decision type confidence — Am I right about what kind of decision this feeds?<br>3. Consequence level confidence — Am I right about how reversible this is?<br>4. Operating posture confidence — Am I right about the conditions?<br>5. Trust level confidence — Has this agent earned the trust level I'm granting?

The formula: routing_confidence = Σ(wᵢ × cᵢ) where weights wᵢ are computed from the work's own characteristics (high-consequence work weights consequence-level confidence heavily; creative work weights classification confidence; delegated work weights trust-level confidence).

Architectural symmetry with deviation measurement. Routing confidence and decision deviation share the same formula applied at different lifecycle points. Deviation measures Σ(wᵢ × dᵢ) retrospectively (how far did we land from expectation?); routing confidence measures Σ(wᵢ × cᵢ) prospectively (how sure are we about where we expect to land?). Both measure the gap between model and reality — differing only in temporal orientation. This symmetry implies they should share the same dimensional structure and weighting mechanism.

Expertise-undermining compromise (Pappu et al. 2026, F1) → No authority model for domain-specific expertise weighting → Vroom-Yetton contingency theory (F10).

Sycophancy / excessive agreement (Yao et al. 2025, F4; Pitre et al. 2025, F9) → No external reference point for conformance checking → Plan-based conformance vs. opinion-based critique (F16).

Structural parameters don't help (Wu et al. 2025, F2) → Governance infrastructure upstream of debate is missing → Routing architecture + topology selection (F10, F11).

Model homogeneity limits diversity (Zhang et al. 2025, F3) → No governance reason for heterogeneity → Domain-owned topology + authority delegation (F11).

Unnecessary debate wastes compute (Eo et al. 2025, F5) → Binary debate/no-debate misses sense-making mode → QuickFire mode (F13, F14).

Confidence is uninformative (Zhu et al. 2026, F6) → Single-scalar confidence collapses dimensions → Five-dimensional routing confidence (F20).

Drift through self-correction (Not addressed in MAD literature) → No bounded self-correction mechanism → Patch budget circuit breaker (F17).

No non-work-product interaction mode (Not addressed in MAD literature) → No sense-making mode exists → QuickFire (F14).

Hallucination undetectable without ground truth (Not addressed in MAD literature) → No plan-based conformance infrastructure → Plan-based hallucination redefinition (F16).

Agents as holons (Koestler): Each agent is simultaneously a self-contained whole (own domain expertise, internal reasoning, local confidence) and part of a larger system (team with shared plans, delegation scopes, collective orientation). The self-assertion/integration balance is what the topology manages (F12).

QuickFire as structural coupling (Maturana): Agents interact through projections without directly accessing internal reasoning — operational closure preserved (F14).

Topology as near-decomposable structure (Simon): Within domain-owned collaboration, intra-domain interactions are stronger than inter-domain interactions — the topology creates the near-decomposable structure that makes governed interaction tractable (F11).

Patch budget as adaptive cycle (Holling 2001): r→K→Ω→α at agent work level (F17).

Authority gradient as medium downward causation (Campbell via S16): The authority holder constrains the decision space (boundary condition) without specifying the outcome (operational directive) — "this must comply with standards" not "change the font" (F12).