Agno adds HITL pause

Agno published new guidance on April 23 showing how its agent framework can pause risky actions until a human approves them. (agno.com) The company said its human-in-the-loop controls work at three layers: individual tools, multi-step workflows, and formal admin approvals. It framed the feature for actions like payments, database changes, and customer-facing messages. (agno.com) (docs.agno.com) At the tool level, a developer can mark a function with `requires_confirmation=True`, which pauses execution before the function runs. Agno’s cookbook example shows that pattern on a Hacker News fetch tool, with the run resuming only after the requirement is resolved. (docs.agno.com) (github.com) At the approval layer, Agno adds a separate `@approval` decorator for actions that need an administrator or compliance review. Its docs say the run pauses, writes a pending approval record to a database, and resumes only after an admin marks it approved or rejected. (docs.agno.com) Agno’s docs list two approval modes: a blocking default for critical actions and an audit mode that logs the event without stopping the run. The company’s examples name deletion, payments, and bulk emails as cases for the blocking path. (docs.agno.com) The workflow version pauses an entire step instead of a single tool call. Agno says that state is persisted in a database so the workflow can stop for confirmation, user input, or output review and then continue from the same point. (docs.agno.com) That matters because agent builders are moving from chat demos to software that can call APIs, run queries, and trigger live systems. Agno’s blog argues that once an agent touches a financial transaction, a production database, or a customer message, the question shifts from speed to approval. (agno.com) Agno also draws a boundary between the layers. Its workflow docs say tool-level pauses inside an agent step are not propagated automatically to the workflow, and developers should use workflow-level human review when they need the whole flow to stop cleanly. (docs.agno.com) The company’s pitch is that human review should be built into the execution model, not added after deployment. Its latest docs and examples turn that argument into concrete patterns: pause, persist, approve, then continue. (agno.com) (docs.agno.com)