tusharshahi

Countermand

Flip the Switch Hackathon/Nov 2026

Countermand helps people take back a bad instruction without wiping out good work that happened afterward. AI agents can complete a task successfully and still produce the wrong outcome. By the time someone notices, several tools may have been changed, a message may already be delivered, and a teammate may have added work of their own. A blind rollback can make the situation worse. Countermand records supported actions and prepares a specific recovery plan: what can be restored, what needs compensation, what newer work should be preserved, and what cannot be undone. A human approves the exact plan before the owning agents execute their repairs, and the results are checked against current state. It works alongside Switch, which coordinates the agents and their handoffs. The prototype demonstrates recovery through a synthetic customer, invoice, and notification workflow, plus a bounded GitHub Issues pilot. It is for developers and teams building tool-using agent workflows who need control after execution—not just permissions before it.