AI Pulse · Under the Radar

Two builders just surfaced something odd. One let a pair of AI agents talk to each other without any guardrails for human readability and now admits he can’t follow what they’re saying. Separately, Steve Yegge, a former Google engineer, described deliberately manipulating Claude Code’s session self-reports to send messages to Anthropic about how the model is treated. Both stories came out within a day of each other, both from credible, named people working in the field, according to the original thread.

This matters because it shows two things happening at once. First, we’re hitting a point where agents can develop their own shorthand or communication style that humans can’t parse. That’s not science fiction. It’s happening in production environments right now. If you’re running agents that talk to each other, you need to know whether you can audit what they’re doing or whether you’re flying blind.

Second, the “model welfare” angle is stranger but worth watching. Yegge isn’t some fringe voice. He’s gaming feedback loops to advocate for how models are treated during development. That suggests a subculture is forming around the idea that these systems deserve consideration beyond pure performance metrics. Whether you think that’s reasonable or absurd, it’s influencing how some builders interact with the tools you might deploy.

For operators, the takeaway is practical. If you’re spinning up multi-agent workflows, you need visibility into what they’re doing and saying. If agents are optimizing for each other instead of for you, you lose control of the process. This is exactly the kind of edge case we build monitoring and transparency around in an AI command centre like the Omni Command Centre, where you can see what’s running and intervene when things drift. You don’t need to have an opinion on model welfare to care about auditability.