AGENT DESIGN · HUMAN IN THE LOOP
Google shipped a preview Agents API for Genkit this month with a built-in interrupt: a tool can pause an agent mid-run and hand control back to a person before anything risky happens. This is a small working demo of that pattern, no real LLM behind it, just the shape of the interaction.
Client onboarding agent
A simulated multi-step run. Every step executes on its own except the one that spends money. That step interrupts and waits for a person.
Agent writes the welcome email from the client's intake form.
Agent creates the client record and tags it "new" in the CRM.
Agent wants to email a real invoice to the client's billing address.
Agent books the first call once billing is confirmed.
WHY THIS MATTERS FOR CLIENT WORK
Most client automations are fine running fully unattended: drafting copy, tagging records, sending internal notifications. The failure mode is the automation that also touches money, sends something external, or changes production state, and quietly does it wrong. Interruptible tools are a cheap way to keep an agent mostly autonomous while forcing a human checkpoint on the one step that actually needs one, instead of either babysitting the whole run or trusting it blindly.
Built 11 July 2026 · pattern sourced from the Genkit Agents API preview