How yutao1024/entcollabbench works
EntCollabBench benchmarks enterprise collaborative-agent workflows, including MCP tool use, role-specialized peer collaboration, and policy-oriented approvals. In the main runtime path, a task reaches an agent, which uses role-selected tools to execute local or Arena MCP operations; approval agents can consult mounted reference documents. The repository also includes a judge component for benchmark scoring. The excerpts do not show the benchmark runner or full agent dispatch implementation, so those connections are not inferred; the graph reflects documented capabilities and sampled module responsibilities.
Main entry points: agent.py Agent service · mcp_bridge.py MCP bridge · toolsets.py Role toolsets · tool_registry.py Tool registry
Components 17
Agent runtime
- Agent service
agent.py - Agent API contract
api_schema.py - Session bindings
session_store.py
Tool execution
- Role toolsets
tool/toolsets.py - Tool registry
tool/tool_registry.py - LangChain tools
tool/langchain_tools.py - Tool executor
tool/tool_executor.py - MCP bridge
tool/mcp_bridge.py - Gitea isolation
tool/gitea_isolation.py
Arena services
- Arena MCP services
Arena - Business seed data
Arena/seed
Approval and evaluation
- Approval references
local_data - Benchmark judge
scripts/judge/judge.py
Other components
- Task requester
- LLM service
- Redis
- Benchmark workflow
Connections 16
- Task requester → Agent service: submits task
- Agent service → Agent API contract: validates requests
- Agent service → LLM service: requests inference
- Agent service → Role toolsets: selects tools
- Role toolsets → Tool registry: resolves keys
- Tool registry → LangChain tools: loads tools
- LangChain tools → Tool executor: executes tools
- Tool executor → MCP bridge: dispatches MCP
- MCP bridge → Arena MCP services: calls tools
- MCP bridge → Gitea isolation: manages Gitea
- Arena MCP services → Business seed data: loads seed data
- Agent service → Session bindings: stores bindings
- Redis → Session bindings: persists bindings
- Approval references → Agent service: supplies references
- Benchmark workflow → Benchmark judge: requests scoring
- Benchmark judge → LLM service: requests evaluation