✓ Accepted KDD 2026 · Workshop on Agentic AI Evaluation & Trustworthiness

How much coordination gain is real?

A paired noise-floor protocol for multi-agent LLM benchmarks. Two protocols whose API inputs are configuration-equivalent — verified by code inspection plus SHA-256 byte audits — can still disagree by up to +18 percentage points on the same benchmark, model, and seed. Seven of ten recent multi-agent coordination architectures report headline gains below that envelope.

Alibek Kaliyev
THE UNIVERSITY OF TEXAS AT AUSTIN
Artem Maryanskyy
UBER

Published headline effects from ten recent multi-agent coordination papers, plotted against the paired noise envelope we measure on τ²-bench retail / Claude Haiku 4.5 (pooled clean-contrast upper CI ≈ 12 pp; largest pooled upper CI ≈ 15 pp). Hover the dots. Red band = indistinguishable from run-to-run noise in our setting.

§1The problem

Small deltas, big claims

Multi-agent coordination papers report benchmark gains of a few points as evidence that one architecture beats another. Almost nobody asks the prior question: how much do two runs that shouldn't differ, differ?

Two LLM agents work in parallel on the same multi-step itinerary. The first attempts to book a flight, discovers there is no inventory, abandons the path, and moves on. The second — started moments later by the same orchestrator — attempts the same booking through the same API and re-discovers the same dead end. The first agent's failure lived only in its private scratchpad.

This is a structural property of how multi-agent LLM systems coordinate today, and it is measured, not anecdotal: the MAST taxonomy attributes 36.94% of failures across 1,600+ production agent traces to inter-agent coordination breakdowns — the single largest category. Recent architectures address it with coordination channels (structured handoffs, shared memory, server-side state) and report small benchmark deltas as evidence one design beats another.

We set out to run exactly that kind of comparison — and the measurement forced a question we should have asked at the start: what is the run-to-run variance between protocols whose API inputs are equivalent? A single number turns out to be enough to invert the read of much of the recent coordination literature.

The one-sentence version

Before you claim your coordination mechanism added +6 pp, you must know that two mechanisms doing nothing at all can differ by +18 pp on one seed — and by −3 pp on the next.

§2The instrument

Three protocols, one store, matched payloads

ET-MCP — a task-scoped negative-knowledge store conformant with MCP 2026-07-28 — is the substrate that makes the measurement possible, not the contribution. Agents write typed failure events; two reader-side designs decide who reads them, and when.

All runs use τ²-bench retail (state-validated passk on a customer-service tool-use domain), with Claude Haiku 4.5 as both assistant and user simulator, temperature 0, a 20-turn budget, n = 100 tasks × 2 trials per protocol. When any trial ends with reward < 1, the final K = 3 tool calls are written to the task-scoped store as failed_trial_action events. The three protocols differ only in the reader interface:

no_coord
baseline · nobody reads

The store is written but never read. Pure baseline: whatever this arm scores is what the benchmark plus harness plus API produce with zero coordination.

pull
agent as reader

The agent is the active reader: a curated peer-warnings summary is prepended to its system prompt at trial start. The agent pays decision overhead and an enlarged tool surface for control.

intercept
framework as reader

The framework intercepts every tool call; if (name, arguments) match a peer failure event, it prepends a [PEER-WARNING] block to the tool response. No agent changes, no prompt changes — the same hook surface already shipped in Strands, AgentCore Policy, LangChain middleware, and Microsoft Agent Framework.

Reader-side architectures
Same store, writer policy, and MCP transport — the designs differ only in who reads and when the payload is presented.
P1Store → agent: curated summary of peer failures, prepended to the system prompt at trial start.
P2Agent → tool: the agent makes its tool call, optionally querying trace.query when peer state might matter.
P3Tool → agent: the raw response returns unmodified. The payload was presented once, up front.
I1Agent → framework: the tool call enters the orchestrator's hook surface.
I2Framework → tool: the call is forwarded unchanged.
I3Tool → framework: the raw response comes back.
I4Framework → store: consult the task-scoped store for a matching peer failure.
I5Framework → agent: on an argument match, a [PEER-WARNING] block is prepended to the response.
At trial 0 the store is empty under every protocol, so both readers are logical no-ops: pull's augmenter returns its input unchanged, and the intercept hook is pass-through. no_coord vs. intercept is the genuinely configuration-equivalent contrast — verified by code inspection plus two SHA-256 audits (40/40 first-request payload cells byte-identical; 30/30 live wire-byte opens identical).
§3Result 1 · The head-to-head

Coordination didn't move the metric

At trial 1 — where the store is populated and coordination is logically active — all three protocols collapse to the same success rate: 0.54. And where the mechanism had a real chance to help, it directionally hurt.

Trial-conditional success · τ²-bench retail, Haiku 4.5, seed 1, n=100
Solid = trial 0 (empty store, coordination inert) · hatched = trial 1 (store filled, coordination active). Red band = pooled noise envelope. Hover bars for Wilson CIs.
Trial-0 differences occur where the mechanisms are inert — that spread is the noise floor, not architecture. Trial-1 rates are flat at 0.54/0.54/0.54 (seed 2: 0.52/0.54/0.51). Unconditional paired trial-1 sign tests over all 100 tasks: p = 1.0 for all three contrasts.

The coordination-active subset: no detectable effect, mild interference

The pre-specified architectural test conditions on tasks where trial 0 failed — there, the store necessarily contains peer events, so the mechanism is active. Pull and intercept tie exactly, and each directionally underperforms no_coord at recovering from a prior failure. With 8–17 informative pairs after ties, the test is underpowered for a medium effect — this is no detectable effect at the available power, not a clean null.

Coordination-active paired sign tests (trial 1, both-failed-trial-0 subsets)
Contrastsubset nt1 successW / L / Tsign p
pull vs. intercept290.28 / 0.285 / 5 / 191.00
pull vs. no_coord280.14 / 0.292 / 6 / 200.29
intercept vs. no_coord350.20 / 0.346 / 11 / 180.33
Each contrast uses its own paired subset; rows are not cross-comparable. Zero infrastructure errors across all 600 trials — a ~700-line Anthropic-native harness bypasses a documented litellm bug class that corrupts ≥30% of Haiku 4.5 trials in the stock pipeline.
§4Result 2 · The signature measurement

The noise floor

At trial 0 every reader is a no-op — any paired gap between protocols is noise, not architecture. We measured that gap across two seeds, two models, and two domains.

+18 pp
largest single-seed paired gap (pull − intercept, seed 1, pcorr=0.012)
−3 pp
the same contrast at seed 2 — sign flipped, pcorr=1.0. Retracted.
[−2,+12] pp
pooled Wilson CI, clean configuration-equivalent contrast (n=200)
≲15 pp
largest pooled upper Wilson CI across any contrast — the envelope
Trial-0 signed paired gaps by seed, model, and domain
Bars = signed paired-sign-test gap in pp. Red band = pooled clean-contrast CI [−2,+12] pp over the two Haiku retail seeds; dotted line = largest pooled upper CI ≈ 15 pp. ★ = the single-seed result that did not replicate. Hover for n and p.
The floor is local, not universal: Haiku airline collapses to ≈0 pp on all three contrasts, while Sonnet retail spans magnitudes 3.3–20 pp with the direction flipped — pull starts trial-0 behind no_coord on Sonnet. Transporting the floor to a new (model, domain) pair means re-running the paired protocol.

The protocol caught its own artifact

At seed 1, pull-vs-intercept showed a +18 pp gap with Bonferroni-corrected p = 0.012 — the kind of number that gets published as a finding. We re-ran the full n = 100 sweep at a second seed: the gap became −3 pp, pcorr = 1.0. The sign flipped. We retracted the hypothesized pull-specific perturbation (E4) as an established source. Pooled over both seeds, no trial-0 contrast is significant after Bonferroni — at either seed or pooled.

What the trial-0 gap decomposes into: E1 byte-identical first requests (audited, holds); E2 within-protocol API stochasticity, bounded at ≤3 pp by a paired n=30 replicate — the Messages API exposes no seed, so T=0 is not bit-deterministic; E3 harness drift once the user-simulator emits its first sampled token; E4 the retracted structural perturbation. The envelope any coordination claim must clear is E1–E3 combined.

A worked example of self-correction

A single seed's pcorr = 0.012 on a contrast that pools to +7.5 pp (pcorr = 0.316) over two seeds is exactly the artifact a paired noise-floor protocol should catch — and ours did, on our own result.

§5Result 3 · The recontextualization

Where published gains sit relative to the floor

Ten recent multi-agent coordination papers, each reporting a single-number headline delta. Seven fall below the clean-contrast pooled upper CI; one more sits inside the envelope.

Published headline effects vs. the Haiku-retail pooled envelope
Click a row to highlight it in the chart. ↓ below 12 pp · = inside 12–15 pp · ↑ above 15 pp.
PaperBenchmark / taskReported effectvs. floor
Motivation, not adjudication: the comparison is heterogeneous (different model, task, metric, n) and any individual row may be entirely real in its original setting — but whether it survives a same-model paired replication is, by construction, untested. That is the point. †CA-MCP reports BERTScore points, not pp.
§6The fix

Coordination-active passk

The minimum reporting protocol: passk restricted to trials where the coordination mechanism is logically active — the store is non-empty, so the mechanism had an opportunity to act.

Marginal passk answers “should I deploy the whole system?” Coordination-active passk answers “did coordination help when it had the chance?” The two diverge whenever empty-store trials carry unrelated configuration perturbations — which is exactly the situation we measure. The metric is causally pre-specifiable from protocol semantics: when the store is empty the mechanism is inactive, so any cross-protocol comparison on empty-store trials measures noise.

The Sonnet probe shows why this matters. At n = 30, pull posts a +23 pp marginal gain — but it started trial 0 at −16.7 pp behind the baseline, so much of that “gain” is recovery of an unrelated deficit. Marginal pass1 obscures the direction flip; the coordination-active metric makes it visible.

Release gate · minimum sample size to detect a claimed effect Δ
τ²-bench retail, α = 0.05, power 0.80, baseline p₁ = 0.50. Paired column calibrated on the measured trial-0 discordance rate (p_d = 0.32). Pick your claimed effect:
tasks per arm · independent two-proportion z-test
tasks · paired McNemar test
full 114-task τ²-retail seeds required (indep / paired)
Read-out: τ²-bench retail's public 114 tasks at one seed can detect only Δ ≥ 20 pp with independent arms, or Δ ≥ 15 pp paired. Effects the size most coordination papers report (+1 to +11 pp) need hundreds to thousands of paired tasks to separate from the floor.

Runtime monitoring, shipped

Three alarms operationalize the metric on AgentCore / SageMaker Model Monitor deployments: (R1) CUSUM on coordination-active passk over a golden-task set, alarm at ±1.5σfloor; (R2) rolling-N 95th-percentile injection-event count, alarm at ≥50% above release baseline (right-skew defeats Gaussian 2σ); (R3) coordination-key collision rate, alarm at >0.5%.

§7Diagnosis

Why naive coordination fails — and what to try next

Three candidate failure modes surface in the trial-level data, each matched to a refinement on hook surfaces that already exist in production frameworks. All three are preliminary — hypotheses for confirmatory work.

M1 · Failure mode

Writer mis-attribution

The causally relevant call is often upstream; a last-K writer records downstream symptoms instead of the root cause.

LLM-judge audit: 77% of 30 analyzable failures rated upstream (single judge — hypothesis only)
P1 · Refinement

Causally-attributed writers

An LLM judge consumes the trajectory and reward and re-weights which calls get written, replacing the recency-only heuristic.

M2 · Failure mode

Per-match injection noise

The store accumulates events (mean 2.77 per trial, max 8) and every matching call fires a separate warning — repeated interruptions, not curated signal.

P2 · Refinement

Selective intercept

A score gate fires only when a peer event is relevant, novel, and decision-proximal. Preliminary: +15.8 pp over vanilla intercept on the coordination-active subset (n=30, p=0.18 — directionally consistent, underpowered).

M3 · Failure mode

Brittle literal keying

Literal argument tuples don't transfer: cancel_order(#W123) doesn't match cancel_order(#W456) even when both target shipped orders.

P3 · Refinement

Predicate constraint extraction

An extractor turns the trajectory into a typed predicate — tool=cancel_order with order.status=shipped — so intercept fires on the predicate, not the literal tuple.

P2 preliminary signal · trial-1 success, coordination-active subset
Haiku 4.5 retail n=30, paired subset n=19, sign test 2/7/21, p=0.18. Hyperparameters not pre-registered; Sonnet at n≥250 is the pre-registered confirmatory test.
§8Scope

What this paper does not claim

The negative reading applies to naive coordination under a measurement-honest protocol — not to coordination as a research program. Six limitations bound the claims.

L1Configuration-equivalent, not identity-replay
Request equivalence is verified by code inspection plus a 40/40 SHA-256 payload-hash audit on a separate n=10 sample; the headline n=100 sweep was not itself byte-audited. The Messages API exposes no seed, so the pooled envelope upper-bounds the residual server-side stochasticity floor rather than isolating it.
L2Writer was pilot-tuned
The K=3 last-calls writer and single event type were tuned on a prior n=30 sweep. Coordination-active passk is pre-specifiable in principle but was applied here to already-frozen data.
L3Two seeds, one within-protocol replicate
A paired n=30 within-protocol audit bounds pure API stochasticity at ≤3 pp. The second n=100 seed confirms no single-seed paired contrast survives: pooled, all three trial-0 contrasts land at 2.5–7.5 pp signed gaps with Wilson upper CIs ≤15 pp and Bonferroni pcorr > 0.3.
L4Single-judge M1 audit
The 77% upstream rate uses a same-model-class judge with no Cohen's κ or independent replication. M1 is a hypothesis for confirmatory work, not a confirmed effect.
L5Cross-model / cross-domain probes are n=30
Sonnet retail and Haiku airline demonstrate regime-specificity of the floor, not coordination claims. Sonnet n≥250 retail is pre-registered as the confirmatory test. The n=100 coordination-active subsets are tight (8–17 informative pairs), so the head-to-head is an underpowered non-result — weaker evidence than the measurement contribution.
L6Reader-side flows only
The protocol assumes peer coordination flows reader-side; writer-side anticipation effects are out of scope. A shared trace store also opens security surfaces (poisoned writes, cross-principal exfiltration) — the prototype assumes a trusted single-tenant deployment.
§9Paper

Read, reproduce, cite

All 600 trials are logged as JSONL artifacts; analysis is reproducible via the analysis CLI in the repository. The harness and ET-MCP substrate ship together, open source.

The core insight here is simple and compelling: before claiming coordination helps, you need to know how much variance exists between runs that shouldn't differ. Nobody in the multi-agent coordination space has done this measurement, and it's overdue. […] especially the self-retraction of the seed-1 p=0.012 result when it didn't reproduce — that last bit is rare and commendable.

— Official Review, KDD 2026 Workshop on Agentic AI Evaluation and Trustworthiness (accepted)
@inproceedings{kaliyev2026noisefloor,
  title     = {How Much Coordination Gain Is Real? A Paired Noise-Floor
               Protocol for Multi-Agent {LLM} Benchmarks},
  author    = {Kaliyev, Alibek and Maryanskyy, Artem},
  booktitle = {KDD 2026 Workshop on Agentic AI Evaluation
               and Trustworthiness},
  year      = {2026},
  url       = {https://abekek.github.io/coordination-noise-floor-protocol/}
}