swarm-test

Most multi-agent failures don’t live inside a single agent — they live in the connections. A dependency that looks fine in isolation becomes a single point of failure the moment three downstream agents rely on it. swarm-test addresses this by performing static analysis on your agent graph: it reads your topology definition, computes reliability propagation, and flags cascade risks and structural vulnerabilities. No live API calls are made, which means analysis runs fast and costs nothing per execution. The core workflow is a pip install, a YAML or code-defined graph, and a CLI invocation that produces a report.

The differentiating feature is the math it applies to graph structure. The project README demonstrates the core problem with a concrete calculation: fourteen agents at 95% individual reliability produce a system-level reliability of roughly 49%. swarm-test operationalizes that reasoning across your actual topology, identifying which nodes or edges are responsible for the steepest reliability drops — information that is invisible when you test agents in isolation.

The tool fits cleanly into CI/CD pipelines — the repo ships an action.yml for GitHub Actions integration, so structural checks can block a merge before a fragile graph reaches staging. Where it breaks: swarm-test is a static analyzer. It models the topology you define, not the runtime behavior of the models inside each agent. Prompt failures, latency spikes, model version drift, and failures that emerge only under concurrent load are outside its scope. Teams that need runtime reliability guarantees use this as a pre-flight check and pair it with live integration tests or an observability layer.

The project is MIT-licensed and self-hostable, with no API dependency and no external service calls required. Report output includes Mermaid graph visualizations, with example outputs committed directly to the repository for reference before adoption.