| name | agent-workflow-graph |
| description | Design and validate graph/state-machine agent workflows with explicit starts, terminal states, guarded cycles, human approval points, and side-effecting tool boundaries. Use when building LangGraph-style, multi-agent, or durable agent workflows.
|
| version | 1.0.0 |
| author | Zijian Guo |
| license | MIT |
| platforms | ["linux","macos"] |
| prerequisites | {"commands":["python3"]} |
| metadata | {"hermes":{"tags":["agent","workflow","graph","state-machine","orchestration","multi-agent"]}} |
agent-workflow-graph - Agent workflow validation
Use this skill when an agent should be modeled as a controlled workflow instead
of an unconstrained loop.
Research Signal
Public agent frameworks are moving toward graph/state-machine orchestration for
durable, inspectable workflows. The gap in many projects is not lack of an LLM,
but missing control-flow invariants: unreachable nodes, unbounded loops, missing
terminal states, and side-effecting tools without approval boundaries.
Workflow
- Describe the workflow as nodes, edges, start, and terminals.
- Mark side-effecting tool nodes and human approval nodes explicitly.
- Check graph reachability and terminal reachability.
- Add guards or iteration limits to cycles.
- Keep the graph spec near the code that implements it.
Script
~/.hermes/skills/agent-workflow-graph/scripts/workflow_graph_check.py
Spec Shape
{
"start": "plan",
"terminal": ["done"],
"nodes": [{"id": "plan"}, {"id": "send_email", "side_effect": true, "requires_approval": true}, {"id": "done"}],
"edges": [{"from": "plan", "to": "send_email"}, {"from": "send_email", "to": "done"}]
}
Pitfalls
- Graph validation does not prove the node implementation is correct.
- Approval nodes should show raw tool arguments, not only the agent summary.
