Experience as a Compass: Multi-Agent RAG with Evolving Orchestration (April 2026)

Free

HERA: 3-layer hierarchical framework that jointly evolves global orchestration strategies and local agent behaviors using experiential knowledge — role-aware prompt optimization drives targeted improvements for each agent's responsibilities

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About Experience as a Compass: Multi-Agent RAG with Evolving Orchestration (April 2026)

HERA is a hierarchical multi-agent Retrieval-Augmented Generation (RAG) framework introduced in the research paper 'Experience as a Compass: Multi-agent RAG with Evolving Orchestration and Agent Prompts.' It addresses limitations of static agent behaviors and fixed orchestration strategies by jointly evolving global orchestration strategies and local agent prompts through experiential knowledge. At the global level, HERA optimizes query-specific agent topologies via reward-guided sampling and experience accumulation. At the local level, Role-Aware Prompt Evolution (RAPE) refines individual agent behaviors using credit assignment and dual-axes adaptation along operational and behavioral principles. Evaluated on six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization, where sparse exploration yields compact, high-utility multi-agent networks.

Key Features

Hierarchical framework with global and local levels

Global level: reward-guided sampling and experience accumulation for query-specific agent topologies

Local level: Role-Aware Prompt Evolution (RAPE) via credit assignment and dual-axes adaptation (operational and behavioral principles)

Joint evolution of orchestration strategies and role-specific agent prompts

38.69% average improvement over baselines on six knowledge-intensive benchmarks

Robust generalization and token efficiency

Emergent self-organization: sparse exploration yields compact, high-utility multi-agent networks

Supports hard queries requiring multiple steps, sources, or complex reasoning

Pros & Cons

Pros

Adapts orchestration dynamically to query complexity
Learns individual agent behaviors from experience, not static prompts
Significant performance gains over fixed multi-agent systems
Token-efficient, reducing unnecessary computation
Generalizes well across different knowledge-intensive tasks
Demonstrates emergent self-organization of agent networks

Cons

Complexity of implementing the hierarchical learning framework
Requires multiple agents, increasing computational overhead
Dependent on reward signals for reinforcement learning components
Validated only on benchmark datasets; real-world deployment not addressed
May require careful tuning of credit assignment and adaptation parameters

Best For

Multi-hop question answering over multiple documentsComplex reasoning tasks that span multiple knowledge sourcesKnowledge-intensive tasks requiring adaptive agent collaborationResearch in multi-agent systems and retrieval-augmented generationBenchmarking adaptive orchestration in AI agent systems

FAQ

What is HERA?

HERA is a hierarchical multi-agent RAG framework that jointly evolves global orchestration strategies and local agent prompts using experiential knowledge, as described in the paper 'Experience as a Compass: Multi-agent RAG with Evolving Orchestration and Agent Prompts.'

How does HERA differ from existing multi-agent RAG approaches?

Existing approaches rely on static agent behaviors and fixed orchestration strategies. HERA introduces continuous adaptation at both global (orchestration topology) and local (agent prompts) levels through reward-guided sampling and role-aware prompt evolution.

What benchmarks were used to evaluate HERA?

HERA was evaluated on six knowledge-intensive benchmarks, achieving an average improvement of 38.69% over recent baselines while maintaining robust generalization and token efficiency.

Is HERA open source?

The paper is freely available on arXiv under a permissive license. Code availability is not explicitly stated on the provided page, but the paper may include links to code repositories.