Tollgate AI: An Evidence-Gated Autonomy Layer for AI Operations Agents

AI agents are increasingly granted tool calling access to production cloud infrastructure, able to restart services, modify configuration, scale resources, and stop or terminate compute instances. No standard mechanism currently enforces a checkpoint between an agent’s decision to act and that action’s actual execution. Recent published research has proposed the concept of gating an AI operations agent’s autonomy behind named evidence and tiered permission levels, but, to the best of our knowledge, no public working implementation of this concept exists. This paper presents Tollgate AI, an open-source gateway that wraps an AI agent’s AWS SDK calls, enforcing three autonomy tiers and requiring explicit, checkable evidence - a cited log entry, a crossed metric threshold, a matched runbook, or a human approval token – before any mutating action is permitted to execute. We describe the design, validate it end-to-end against a live AWS account, and report a controlled before/after evaluation across 20 synthetic agent-action scenarios, in which the gate reduced bad-outcome actions from 8/20 to 3/20. We report this result alongside its limits with equal weight: the gate verifies that evidence exists and is well-formed, not that it is true, and a blast-radius safety cap blocked two scenarios that were in fact legitimate. We position this as a narrow, honestly scoped contribution rather than a general solution and describe what remains open.