Multi-Agent Collaboration Orchestration Engine

Problem

In multi-agent workflows, the hard part is often not a single model's capability, but chaining collection, analysis, and reporting into a stable process: upstream outputs can be lost, a failure may force a full rerun, branch logic needs manual supervision, and execution state is hard to inspect. This skill treats the AI as an orchestrator and uses local scripts to manage a DAG pipeline's validation, scheduling, state, and recovery.

How It Works

Key capabilities include:
- DAG definition and validation: define nodes and dependencies in JSON, then check required fields, cycles, orphan nodes, and output a topological order.
- State sharing: node outputs are written to pipeline_state.json, and downstream nodes read upstream results automatically.
- Error recovery and resume: handle failures with retry → fallback → abort, and resume from the failed node while skipping completed work.
- Reports and visualization: generate Markdown execution reports and an HTML Gantt chart with statuses such as completed, failed, running, skipped, and pending.
- Dynamic control flow: support condition, switch, for-each, and while-loop, allowing runtime decisions based on node outputs instead of a fixed graph.

Boundaries

It is aimed at local multi-step orchestration, not real-time stream processing, distributed multi-node scheduling, or GPU resource scheduling. The actual collection, analysis, and reporting tasks are still performed by the AI; the skill manages the flow, state, and recovery.