Human
Pause execution until someone approves or declines from the inbox. Downstream logic can branch on the decision using mapped output keys.
Overview
The approval step is how you insert human judgement into an automated graph: policy sign-off, risky transfers, publication gates, or any checkpoint that must not run purely on model output.
Steps in this family
Today the human family centres on a single Approvaltemplate: configure copy, timeouts if applicable, and outputs that carry the reviewer's decision into the next nodes.
Configuration
Use General for labels and user-facing messaging where the product exposes it, and Output to map decision metadata (for example decision enum, timestamps) from {{exe.*}} after a response. There is no Execution tab: the work happens in the inbox UI instead of inside a model or sandbox.
Data flow
While waiting, the run sits in a paused state; after action, the runner emits the mapped output object like any other step so branches and AI steps can continue.
Related
- Assistant tools and approvals: product-level context for approvals where documented.
- Expression variables
- Steps and behaviour hub