LN: Gao et al. (2024) — AgentScope: A Flexible yet Robust Multi-Agent Platform

Bibliographic Reference

Citation: Gao, D., Li, Z., Pan, X., et al. (2024). AgentScope: A flexible yet robust multi-agent platform. arXiv:2402.14034. https://arxiv.org/abs/2402.14034 Affiliation: Alibaba Group


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategoryFramework + system proposal
ContextAlibaba’s answer to AutoGen/LangGraph; platform for multi-agent applications
CorrectnessEvaluated on diverse multi-agent tasks including coding and QA. Robustness tests.
Contributions(1) Actor-model concurrency for parallel agent execution; (2) Fault-tolerant pipeline with automatic retry; (3) Multiple LLM backends (OpenAI, Ollama, Hugging Face); (4) Built-in monitoring and debugging
ClarityExcellent. Comprehensive documentation.

Relevance: ⭐⭐⭐⭐

AgentScope explicitly supports Ollama as a backend, making it directly compatible with PUMA’s local inference stack.


Pass 2 — Key Points

AgentScope’s actor-model concurrency allows running multiple specialist agents in parallel — directly applicable to PUMA Stage 5 where a triage agent, estimation agent, and risk agent could process a backlog simultaneously.

Ollama integration: AgentScope’s OllamaChat backend calls Ollama’s REST API. This means PUMA could switch from a custom Ollama integration to AgentScope with minimal code changes.

Comparison with LangGraph: AgentScope excels at robustness and concurrency; LangGraph at graph-based state machines with cycles. For PUMA Stage 5, AgentScope may be simpler to implement correctly.


PUMA Integration

MOCs