Cloudflare published a post-mortem on Project Glasswing, an internal initiative that put an AI agent named Mythos to work on real tasks inside the company. The writeup covers what the experiment exposed about how AI agents behave in practice versus expectation.
Mythos - Cloudflare's name for the agent they built and tested under the project - ran into the kinds of problems that consistently trip up production AI systems: task ambiguity, unexpected failure modes when chaining actions together, and the gap between what an agent can reason through and what it can reliably execute. These aren't hypothetical concerns. Every company deploying agents in real workflows is hitting versions of the same walls.
Cloudflare is worth paying attention to here because they sit at the infrastructure layer of the internet, not at the application layer. Their observations about agent reliability carry different weight than a startup's blog post - they're seeing what happens when autonomous AI processes touch systems that handle millions of requests. The specifics of what Glasswing surfaced, particularly around agent decision-making at scale, are the kind of ground-truth data points the field still badly needs more of.