Self-Evolving Hybrid Agentic Workflow DAGs
MAP-Elites evolution of LLM + Code node topologies for autonomous app development. Based on arXiv:2603.19639.
HyEvo treats the workflow topology itself as an evolvable artifact. Instead of fixed agent pipelines, it generates directed acyclic graphs (DAGs) mixing two node types:
- LLM nodes -- semantic reasoning (code analysis, planning, review)
- Code nodes -- deterministic execution (build, test, lint)
A multi-island MAP-Elites algorithm evolves these topologies over generations, discovering optimal structures per app.
Seed Population (3 topologies x 6 islands)
|
v
Execute Best DAG -----> Score Fitness
^ |
| v
Migrate <---- Evolve <---- Reflect
(every Nth) (MAP-Elites) (LLM meta-agent)
Fitness: R = 0.9 * quality + 0.05 * cost_utility + 0.05 * latency_utility
6 Islands, each with a mutation strategy:
| Island | Strategy | Effect |
|---|---|---|
| 0 | Add Node | Insert LLM or Code node |
| 1 | Remove Node | Prune low-impact node |
| 2 | Swap Type | LLM <-> Code conversion |
| 3 | Rewire Edge | Change connections |
| 4 | Mutate Content | LLM-refined prompts/scripts |
| 5 | Crossover | Merge two parent DAGs |
Parent selection: 50% elite, 30% history, 20% cross-island.
Sources/HyEvo/
Models/
WorkflowDAG.swift # DAG, Node, Edge models + seed topologies
WorkflowPopulation.swift # MAP-Elites population, islands, mutations
AppEvolutionState.swift # Pipeline state machine + HyEvo stages
Views/
WorkflowDAGView.swift # Force-directed DAG visualization (SwiftUI)
FleetEvolutionView.swift # Fleet dashboard with HyEvo integration
Services/
FleetEvolutionClient.swift # Evolution engine, DAG execution, reflect-then-generate
docs/
HYEVO.md # Detailed technical documentation
supabase/migrations/
20260325_hyevo_columns.sql # Database schema (Supabase/Postgres)
audio/
HyEvo_Explainer.mp3 # Audio walkthrough (OpenAI TTS)
- Hybrid nodes: LLM for reasoning, Code for deterministic ops. Evolution finds the optimal ratio (paper sweet spot: 30-50% code).
- MAP-Elites: Quality-diversity search over a 3D feature grid (LLM ratio, depth, edge density). Maintains diverse topologies instead of converging on one.
- Reflect-then-generate: LLM meta-agent analyzes execution feedback, then mutations are applied informed by that analysis.
- Ring migration: Elite DAGs spread across islands every N generations.
Built for MeshControl, an iOS command center for distributed autonomous app development across a multi-machine mesh. HyEvo replaces fixed evolution pipelines with self-evolving workflow topologies.
Adapted from the paper:
HyEvo: Self-Evolving Hybrid Agentic Workflows for Efficient Reasoning Beibei Xu, Yutong Ye, Chuyun Shen, Yingbo Zhou, Cheng Chen, Mingsong Chen East China Normal University, Beihang University, SUIBE, Fudan University arXiv:2603.19639 (March 2026)
Key results from the paper: up to 19x cost reduction and 16x latency reduction vs state-of-the-art baselines on math/coding benchmarks.
MIT