The AI Agent Control Layer Nobody Talks About

A lot of agent control discussion still sits at the wrong layer. Observability tells you what happened. Guardrails help shape behavior. Neither answers the production question that matters most when agents are looping, retrying, fanning out: Can this agent still act — given what it has already done? That's the control point I've been building toward with Cycles. Simple example — a support agent with CRM and email access: Without a runtime decision layer: the customer email fires (may be multiple emails). With Cycles: blocked before execution. The function never runs, emails don't go out. Emails that fire are customer commitments, a compliance exposure, or a support promise. Demo here: https://github.com/runcycles/cycles-agent-action-authority-demo