agentscope-ai / agentscope-ai/agentscope
GNAP: git-native message passing for AgentScope distributed deployments
- Linguagem predominante
- Python
- Estrelas
- 31.6k
- Forks
- 3.5k
- Merge médio
- 1d 16h
- PRs com merge (30d)
- 103
Descrição
Hi AgentScope team 👋
AgentScope has impressive infrastructure for distributed multi-agent systems (the `DistributedAgent` with RPC backend). I wanted to propose a complementary, zero-infrastructure coordination alternative worth considering.
**[GNAP](https://github.com/farol-team/gnap)** (Git-Native Agent Protocol) uses git as the message and task bus. Agents read `.gnap/tasks/` and `.gnap/messages/`, do work, commit and push. No gRPC servers, no service mesh.
**AgentScope-specific angle:**
AgentScope's distributed mode requires setting up `RpcAgentServerLauncher` and managing gRPC connections. This works well for high-throughput scenarios but adds operational overhead. GNAP offers a simpler option for scenarios where:
- Task latency is measured in seconds/minutes (not milliseconds)
- Agents run on heterogeneous infrastructure (local machines, cloud VMs, CI workers)
- You want a full audit trail by default (git history)
- Human collaborators need to participate in the agent pipeline
**Comparison for AgentScope distributed scenarios:**
| | Setup | Infrastructure | Audit trail | Human-in-loop |
|---|---|---|---|---|
| AgentScope RPC | gRPC servers | Required | No | Complex |
| **GNAP** | `git init` | None | Git history | Any git user |
A `GNAPMsgHub` could replace `MsgHub` for offline/async agent coordination while AgentScope's native RPC handles real-time scenarios.
Would the team be open to a GNAP backend as a lightweight alternative for async workflows?
https://github.com/farol-team/gnap
Guia de contribuição
Avaliação
Esta issue ainda não foi avaliada.