Back to Ruflo

Performance SOTA Report — 2026-07-25

docs/dream-cycles/2026-07-25-performance-sota.md

3.34.04.2 KB
Original Source

Performance SOTA Report — 2026-07-25

TL;DR: AA-AgentPerf (June 2026, Grade A) establishes 23.6× hardware efficiency gap (61,354 vs 2,594 Agents/MW) and ACL 2026 Pareto-optimal research proves mixture-of-agents beats self-consistency by +2.7pp — Ruflo benchmarks neither.

What's New in 2026

FindingSourceConfidence
AA-AgentPerf: first agentic inference benchmark, Agents/MW metric; GB300 = 61,354 vs H200 = 2,594 (23.6× gap)Artificial Analysis, Jun 12 2026A
Mixture-of-agents Pareto-optimal: +2.7pp over self-consistency on MMLU-Pro; most efficient when parallel > sequentialarXiv 2605.01566, ACL 2026 SRWA
RecursiveMAS: 1.2–2.4× inference speedup, 34.6–75.6% token reduction, +8.3% accuracy across 9 benchmarksarXiv 2604.25917B
AutoAgents-Rust: 4.97 rps vs LangGraph 2.70 rps (+84%); CrewAI 44% failure rate excludeddev.to/saivishwak, Feb 2026B
Mixture-of-agents at 20× CoT budget yields +7.1% MMLU-Pro over chain-of-thought baselinearXiv 2605.01566, ACL 2026 SRWA

Ruflo Current Capability

CapabilityStatusNotes
Agent inference benchmark❌ No AA-AgentPerf registrationperformance benchmark --suite all exists but no Agents/MW metric
Test-time compute scaling❌ AbsentNo mixture-of-agents routing config; 3-tier routing is cost-only
Flash Attention speedup❌ UnverifiedCLAUDE.md: "unverified — no benchmark exists for this claim"
MCP response target⚠️ Unmeasured<100ms target set but no CI gate
Framework throughput⚠️ UnbenchmarkedNo public rps/latency numbers

Competitor Comparison

FrameworkThroughput (rps)P95 Latency (ms)Failure RateTest-Time ScalingAgents/MW
RufloUnbenchmarkedUnbenchmarkedUnknownNoneNot registered
LangGraph2.70 (B)16,891 (B)0%NoneNot registered
AutoGen/AG2~2.0 (C)~15,000 (C)LowSelf-refine (B)Not registered
CrewAIExcluded (B)44% (B)NoneNot registered
OpenAI SwarmLowest latency (C)Low (C)NoneNot registered
AutoAgents-Rust4.97 (B)9,652 (B)0%NoneNot registered

(B = vendor/dev benchmark 2026; C = single source)

Benchmarks

BenchmarkKey NumberGradeReference
AA-AgentPerf GB30061,354 Agents/MWAArtificial Analysis Jun 2026
AA-AgentPerf H2002,594 Agents/MWAArtificial Analysis Jun 2026
Mixture-of-agents MMLU-Pro+7.1% over CoT @ 20× budgetAarXiv 2605.01566, ACL 2026
RecursiveMAS speedup1.2–2.4× end-to-endBarXiv 2604.25917
AutoAgents-Rust vs LangGraph+84% throughputBdev.to Feb 2026

SOTA Proof & Witness

Session commit: 26c35b59b40a0a95b286ccf5ac675a15edcc995f Report SHA-256: 8f7e3d612e5d165f64a2beb23bc314153dc016e1ee479185d63107294c9cd9e5 Witness stamp: e88eb2c46dc489827e399c0a0309590f04e635884ed183d84cac9436106cad45

Verifier: sha256sum report.md → strip newline → concat session commit → sha256sum → must equal Witness stamp.

  1. Register Ruflo in AA-AgentPerf (ADR-320): Add performance benchmark --suite agentperf subcommand that replays agentic trajectories and reports Agents/MW metric. Target ≥5,000 Agents/MW on standard dev hardware as a CI gate. This closes the single largest credibility gap vs verifiable external benchmarks.

  2. Implement mixture-of-agents routing in 3-tier model config (ADR-026): When estimatedParallelAgents > estimatedSequentialAggregations, activate mixture-of-agents path. ACL 2026 Grade A evidence shows +2.7pp over self-consistency at equal compute cost. The hook point is hooks route → tier selection.

  3. Benchmark and gate Flash Attention in CI: Run performance benchmark --suite flash-attention in each PR that touches the attention layer. Ship a Grade A number or remove the unverified claim from CLAUDE.md. Ongoing unverified claim degrades trust in the performance section.