ContextQMD
Back to Ruflo

ADR-333: Efficiency-First Orchestration — Pre-hoc Failure Inference and Adaptive Agent Count Routing

v3/docs/adr/ADR-333-efficiency-first-orchestration-prehoc-failure-inference.md

3.34.04.4 KB
Original Source

ADR-333: Efficiency-First Orchestration — Pre-hoc Failure Inference and Adaptive Agent Count Routing

Status: Proposed
Authors: claude (dream-cycle agent, 2026-07-30)
Date: 2026-07-30
Related issues: dream-cycle 2026-07-30 (#TBD)
Supersedes: none — extends ADR-026 (3-tier model routing) and ADR-322 (provenance typing)

Context

Two Grade-A papers published July 29, 2026 challenge Ruflo's current swarm-invocation policy:

  1. "Two Calls Beat Five Agents" (arXiv 2607.26922): On Qwen2.5-7B, 2-call self-refinement achieves 86.2% GSM8K accuracy at 7.4× lower token usage than a 5-agent pipeline (86.7% accuracy). The accuracy gap is 0.5 pp; the cost gap is 7.4×. For tasks within the capability range of the current model, multi-agent spawning is net-negative on value-per-token.

  2. HalluProp — Pre-hoc Failure Risk Inference (arXiv 2607.26836): Semantic misalignment between agent roles and tasks causes hallucinations to propagate along agent communication chains. HalluProp's pre-task topology risk analysis achieves 84.6% average AUROC for fault localization at 65× the speed of post-hoc methods, sub-second per assessment.

Ruflo's CLAUDE.md currently prescribes "AUTO-INVOKE SWARM when task involves 3+ files" and maxAgents: 8 unconditionally. This policy ignores task-level capability proportionality, leaving 3–7× token efficiency on simple tasks.

ADR-026 added cost-tier routing (deterministic/Haiku/Sonnet). This ADR adds a pre-tier efficiency gate that considers whether spawning agents at all is warranted.

Decision

§1 — Pre-hoc Failure Inference hook in hooks pre-task

Before routing to any agent tier, estimate task propagation risk:

  • Input signals: task description, agent types requested, topology, file count, estimated context size
  • Risk classifier: lightweight rule-based + embedding similarity check (ported from HalluProp topology analysis)
  • Output: { risk_score: 0.0–1.0, recommendation: "self-refine" | "haiku" | "swarm", reasoning: string }
  • Threshold: risk_score < 0.3 AND file_count <= 2 → recommend self-refine
  • Logging: all decisions appended to .swarm/routing-decisions.jsonl for SONA training input

Implementation target: v3/@claude-flow/hooks/src/workers/pre-hoc-inference-worker.ts

§2 — Efficiency routing mode in hooks route

Extend hooks route with --efficiency-mode flag:

  1. Attempt single-model self-refinement (max 2 calls, same model)
  2. If self-refinement score (measured by task completion signal) is below threshold, escalate to Haiku (2 agents max)
  3. If Haiku multi-agent score is below threshold, escalate to Sonnet swarm

Feature gate: CLAUDE_FLOW_FEATURE_EFFICIENCY_ROUTING=1
Dry-run mode: --efficiency-dry-run logs recommendations without applying them.

§3 — Token-efficiency benchmark in performance benchmark

Add --suite efficiency to the performance benchmark command:

  • Metric: Agents/token-budget (tasks completed per 10k tokens spent)
  • Compares: self-refine path vs 2-agent vs 4-agent vs 8-agent hierarchical
  • Integrates with existing AA-AgentPerf gap tracked in ADR-026 (2026-07-25 cycle)
  • Runs on every PR touching hooks/, swarm/, or @claude-flow/cli/src/commands/

Consequences

Positive:

  • 3–7× token reduction on simple tasks (Grade A evidence for 7.4×)
  • Sub-second pre-task risk assessment (HalluProp pattern; 65× speedup vs post-hoc)
  • SONA neural training gains data from routing decisions (aligns with §8 intelligence pipeline)
  • Measurable efficiency metric for CI gating

Negative / risks:

  • Pre-hoc classifier can miss genuinely complex tasks that look simple (false negative → under-spawn)
  • Adds latency to hooks pre-task for every invocation (must stay <50ms to not block CLI)
  • CLAUDE_FLOW_FEATURE_EFFICIENCY_ROUTING=1 gate adds configuration surface

Mitigation:

  • Gate behind env var; off by default for existing users
  • Classifier must time-box at 50ms and fall back to current routing on timeout
  • Log every recommendation mismatch to .swarm/routing-decisions.jsonl for tuning

Alternatives Considered

  • No change: Acceptable for large-scale complex tasks; inefficient for 60–70% of common single-file/simple-question interactions.
  • Lower maxAgents globally: Blunt; doesn't route by task complexity.
  • Increase self-refinement budget in ADR-026 Tier-2: Doesn't address pre-task risk or 2-call path for Tier-1 tasks.