Argus is an open source tool designed to detect silent failures in AI agent workflows before deployment. It integrates primarily with LangGraph based agents and monitors the outputs of individual nodes within an agent graph. When a node returns an empty result, drops a required field, or swallows a tool error without raising an exception, Argus identifies the specific node responsible for the failure. It then halts the build process in continuous integration (CI) environments, preventing the faulty agent from being deployed. The tool operates by attaching to a compiled LangGraph graph using the ArgusRecorder class. It listens to LangGraph’s native callbacks and records the update returned by every node. When a downstream node crashes due to missing data, Argus traces the error back to the upstream node that produced the empty or incorrect output. It provides evidence such as the empty update or missing field. In CI, the command “argus check” exits with a non zero status to signal failure. Developers can then use “argus replay” to rerun the agent from the fixed node, freezing all upstream outputs to avoid redundant and costly LLM calls. Key features include field contracts, which allow developers to declare which fields a node should read and write. A field that is never written, written empty, or dropped triggers a failure on the responsible node rather than on the downstream node that encounters the missing data. The tool also includes “argus fix,” which generates a paste ready prompt for coding agents such as Claude Code or Cursor. This prompt targets the origin of the bug, cites the relevant source line, and operates offline. Argus supports integration with pytest and GitHub Actions for automated testing and deployment pipelines. Typical use cases involve teams building and shipping LangGraph based AI agents who need to catch subtle data flow errors that do not cause immediate crashes. The workflow involves wrapping the compiled graph, running the agent, and checking results in CI. Developers then inspect the root cause, apply a fix, and replay the run from the corrected node. Technical requirements include a LangGraph based agent and a local first setup where user keys remain on the local machine. The core of Argus is open source.