Adding Human Vetoes to Agent Workflows
- •Kondratiuk integrated mandatory human veto gates into his PM agent to prevent inaccurate Slack status updates.
- •Initial 'rubber-stamping' behaviors by the user necessitated adding manual text-entry friction to ensure actual review engagement.
- •Refined workflows now prioritize human gates only for irreversible actions, using a severity-based classification system.
Mykola Kondratiuk, an automation engineer, implemented a mandatory human veto gate for his project management agent after experiencing two embarrassing errors in three months. The agent, which pulls data from Jira to format and post weekly summaries to Slack, previously posted stale information and misidentified overdue tickets. Instead of refining the agent's logic, Kondratiuk integrated a human approval step as a structural component of the workflow, requiring a manual thumbs-up emoji before any team-wide communication occurs. This design follows principles observed in Microsoft Conductor, which treats human intervention as a default requirement rather than an optional safety feature.
Within three weeks, Kondratiuk discovered that simple approval buttons led to 'rubber-stamping,' where the human user stops carefully reviewing drafts, creating an invisible failure mode. He resolved this by introducing mandatory friction: reviewers must now type at least one word to confirm, rather than simply clicking an emoji. This change ensures active human engagement in the decision process.
Further refinements to the workflow included a three-tier severity classifier for decisions. Low-impact actions are now auto-approved, medium-impact tasks receive a soft review prompt, and high-impact actions require a hard gate. Additionally, Kondratiuk reduced the approval timeout from 2 hours to 30 minutes, with an escalating Slack ping if the task remains pending. He established a clear mental model for gate placement: if an action would be difficult to undo at 11pm on a Friday, it requires a human gate. Over 6 weeks of operation, the system successfully blocked two errors and improved draft quality while requiring only 3 minutes of daily overhead.