-
Notifications
You must be signed in to change notification settings - Fork 2
research(orchestration): cascade-aware routing — geometry-switching for tree-like vs cyclic DAGs (arXiv:2603.17112) #2425
Description
Source
arXiv:2603.17112 — Cascade-Aware Multi-Agent Routing: Spatio-Temporal Sidecars and Geometry-Switching (2026-03-17)
Technique
A spatio-temporal sidecar module that runs alongside the routing graph and:
- Measures curvature of the task graph at runtime to classify it as tree-like (hierarchical) or cyclic (looping/iterative)
- Switches routing geometry: Euclidean distance scoring for tree-like tasks, hyperbolic distance for cyclic/iterative tasks
- Detects failure propagation and reroutes around failing cascade paths
Reports success rate improvement from 50% to 87% on multi-hop tool-calling benchmarks.
Applicability to Zeph
TaskGraph DAG (zeph-core/orchestration): The DagScheduler currently uses a static topology with default_failure_strategy = "abort". Geometry-switching could detect when a failing subtask is causing cascades and switch to a fallback routing path.
LlmPlanner: When the planner creates a task graph, it could annotate topology type (tree vs cyclic), letting the sidecar apply the appropriate routing geometry.
Relationship to #2407 (genealogy-graph defense for error cascades): Both papers address cascade failures. This paper's geometry approach is complementary — #2407 builds a diagnosis graph, this paper prevents failure propagation via routing.
Implementation sketch
- Add topology detection in
DagScheduler::execute()— classify DAG as tree/cyclic at plan time - Route tree-like plans with existing priority queue; cyclic plans with cost-weighted retry routing
- Config:
[orchestration] topology_selection = true(already exists — this extends its semantics)
Related
- research(orchestration): error cascades in multi-agent DAGs — genealogy-graph defense raises success rate 0.32→0.89 (arXiv:2603.04474) #2407 — genealogy-graph error cascade defense
- Issue research(orchestration): VeriMAP per-subtask verification predicates — catches handoff mismatches before cascade (arXiv:2510.17109) #2269 — VeriMAP per-subtask verification predicates