wtf is cockpit-core?

Cockpit-core is a Python tool built to watch AI agents while they work and catch problems before they cause real damage. Instead of only checking whether a task finished, it tries to predict where a workflow is heading and step in at the boundary, right before an action that cannot be undone. The idea behind it is that AI agents often fail quietly: they make up a plausible looking input, get back an empty or wrong result, and then report success anyway. Cockpit-core exists to catch that kind of mistake before it happens, not after the fact. The project describes agent failures using a nine state model of how a workflow can drift away from being healthy, either in reversible ways, such as losing context, stalling, drifting from the original goal, or losing track of the truth, or in ways that cannot be undone, such as false completion, runaway behavior, being taken over, or firing an irreversible action. Cockpit-core tracks a workflow's phase, its health, and who currently has control, and combines two detection methods: one that checks for sudden false completions, and one that tracks gradual drift over time using a simple statistical average. The project states that testing on real traces showed these two methods cover most of the failures they measured. Using it is meant to be simple. A developer can add one line to wrap an existing agent function, or use a more detailed set of calls to observe each step, mark phase changes, and gate risky, irreversible actions behind a check that a human can approve or deny. It ships with adapters for the OpenAI Agents SDK and LangGraph, so it can plug into agents built with those frameworks without much extra setup. There is also a local, customer owned option to save records of past workflow runs, with rules for redacting sensitive information such as emails or secrets, so a team can build a history of runs and use it to tune the detectors to their own workload. Cockpit-core comes with a test suite and seven example workflows in the repository, showing situations like a refund agent and an invoice reconciliation task. It is aimed at developers building or maintaining AI agents who want an early warning before a costly mistake ships.