Eino Multi-Agent Notes

中文

CyberStrikeAI uses CloudWeGo Eino ADK for the current single-agent and multi-agent execution paths. The native legacy ReAct path has been removed.

Entrypoints

Single-agent: /api/eino-agent and /api/eino-agent/stream
Multi-agent: /api/multi-agent and /api/multi-agent/stream

Multi-agent orchestration is selected by request body:

deep
plan_execute
supervisor

Robots default to robot_default_agent_mode, and batch tasks can opt into multi-agent through config.

Agent Definitions

Markdown agents live under agents/.

Typical files:

agents/orchestrator.md
agents/orchestrator-plan-execute.md
agents/orchestrator-supervisor.md
agents/*.md

Front matter controls name, id, description, tools, bound role, max iterations, and optional orchestrator kind.

Middleware

Important Eino middleware:

tool search: exposes a small visible tool set and unlocks others on demand;
patch tool calls: repairs interrupted histories;
plan task: structured task board;
reduction: truncates or persists large tool outputs;
summarization: compresses long contexts;
checkpoint: resume after crash/OOM.

These settings live under multi_agent.eino_middleware.

Skills

Eino Skills support progressive disclosure. The Agent initially sees names and descriptions; details are loaded only when needed through the configured skill tool.

Operational Notes

Tool visibility is not the same as tool availability in the UI.
Running streams keep their startup context even if config changes mid-run.
Summarization can write transcripts under data/conversation_artifacts/....
High-risk tools should still be constrained by roles and HITL.

Source Anchors

Multi-agent handler: internal/handler/multi_agent.go
Preparation: internal/handler/multi_agent_prepare.go
Orchestration: internal/multiagent/eino_orchestration.go
Run loop: internal/multiagent/eino_adk_run_loop.go
Skills: internal/multiagent/eino_skills.go
Middleware: internal/multiagent/eino_middleware.go

2.1 KiB Raw Blame History