I am publishing this repository because the incident documented here was my responsibility, and I think other teams can learn from it.
The short version is:
- I ran autonomous workflows too far
- I gave them too much continuity and too little operational restraint
- I let an experimental runtime behave like a production system
The result was costly, embarrassing, and operationally disruptive. It also made one thing very clear: agent systems need stronger guardrails than “it works” or “the code runs.”
I am not publishing raw logs or private artifacts. This repository is intentionally sanitized. The purpose is to make the architecture, the failure modes, and the lessons visible without exposing credentials, identifiers, or unrelated private data.
What I want other builders to take from this:
- do not default to continuous loops
- do not post raw assistant text into public systems
- do not run long stateful workflows without hard budgets and stop rules
- do not assume an AI coding agent will automatically make the safe design choice
The mistake was mine, but the lesson is broadly useful.