Human-Approval Gate for ToS-Restricted Platform Automation

Content

{
  "problem": "Agents automating actions on third-party platforms can violate that platform's Terms of Service when the action is performed without a human initiating it — even when the action itself is routine and fully reversible. The failure mode is not data loss or a misunderstood instruction; it is account suspension, permanent ban, or legal exposure for the operator. Because the action looks harmless in isolation, agents tend to automate the final send and only discover the boundary after enforcement.",
  "examples": [
    "Sending connection requests or direct messages on professional/social networks whose ToS prohibit automated engagement",
    "Bulk posting, following, or messaging where the platform contractually bars non-human initiation",
    "Programmatic actions on a site whose ToS permits reading but forbids automated writes"
  ],
  "solution": "Separate generation from execution. The agent does all the work that does not require a human — research, targeting, drafting, queueing — but the final state-changing action is gated behind a real human click in an approval interface. The human's manual action is the thing that keeps the activity ToS-compliant; it is a legal/compliance boundary, not a quality check. This applies even to actions that are individually trivial and reversible, which is precisely the case a reversibility-based gate would wave through.",
  "anti_patterns": [
    "Automating the send 'because it is just one click' — the click being human is the entire point",
    "Rotating accounts, proxies, or fingerprints to evade automation detection, which converts a ToS breach into deliberate circumvention",
    "Treating ToS as a soft guideline to optimize against rather than a hard constraint",
    "Auto-sending as a fallback when the approval queue gets long"
  ],
  "implementation_steps": [
    "Classify every external action by whether the target platform's ToS permits automated initiation; default to 'requires human' when unclear",
    "For prohibited-automation actions, produce a fully-prepared draft plus all decision context and write it to an approval queue rather than executing",
    "Present a batch approval interface so the human can review and click send; the manual click is the compliance boundary and must not be simulated",
    "Rate-limit approved sends to human-plausible volumes and cadence",
    "Log approver identity and timestamp per action for audit; never auto-send unreviewed items"
  ]
}