Documentation rots. You draw your multi-agent architecture once, then six months later it's wrong because the system changed and nobody updated the picture.
swarm-test v0.3.3 generates the diagram from your actual agent topology, so it stays accurate. And when I ran it on my own 14-agent system, the problem jumped out immediately.
[ATTACH THE AREENGINE DIAGRAM HERE]
That red node is OrchestratorAgent — a single point of failure with 92% blast radius. Look at the shape: nearly every one of the 14 agents funnels into two hubs, OrchestratorAgent and TrainerAgent. If either goes down, the system collapses. You can't see that in a console table. You see it in one glance at the diagram.
Generating it is one command:
swarm-test graph my_crew.py --format mermaid
You get Mermaid syntax to paste straight into a GitHub README:
graph TD
OrchestratorAgent[OrchestratorAgent ⚠️ SPOF]:::spof
TrainerAgent[TrainerAgent]:::healthy
ImageValidatorAgent --> OrchestratorAgent
classDef spof fill:#ff4444,stroke:#cc0000,color:#fff
classDef healthy fill:#44cc44,stroke:#22aa22,color:#fff
GitHub, GitLab, and Notion render this natively. Single points of failure show red, healthy agents green, moderate-risk yellow — the same risk classification from the reliability analysis.
Three formats depending on where the diagram is going:
Mermaid — for READMEs and wikis. Renders inline, version-controllable as text, diffs cleanly in PRs.
DOT — for Graphviz pipelines and custom tooling.
PNG — for slide decks and external docs:
swarm-test graph my_crew.py --format png --output topology.png
(PNG needs matplotlib: pip install swarm-test[png])
The value scales with system size. On a 3-agent crew you can already see the structure in the console. But on a 14-agent system with 40+ edges, the diagram reveals clusters, bottlenecks, and isolated agents that a table can't show. The shape of the problem becomes visible — and the fix becomes obvious. In my case: break the Orchestrator bottleneck by distributing routing across multiple agents.
Because the diagram comes from the same graph analysis that runs the reliability tests, your documentation and your testing never disagree.
Works across CrewAI, LangGraph, AutoGen, and custom orchestrators.
pip install swarm-test --upgrade
GitHub: github.com/surajkumar811/swarm-test
Top comments (0)