You'll noticeTier: how much it should matter to you
4Useful: my rating, 1 to 5
3Signal: worth watching, 1 to 5
WorkflowsArea: what it touches
ImprovementsKind: in v2.1.312,
What probably matters to youSection of the release
What
Workflow agents are subagents, helper agents that Claude starts, that run as steps of a workflow. How they return results changed in two linked ways.
Handback tool: when the agent has no output schema and a permission-mode check passes, its definition is wrapped and given handbackOptIn and a handbackTool. The tool captures the agent's final report.
Final text: the result now comes from the handback message instead of the last thing the agent said.
New result fields: toolReportReview, toolReportUnjudged and toolReportAgentId.
Review: workflow agent output is now reviewed when toolReportReview or toolReportUnjudged is set. Before, the review ran only in auto mode.
Review details: it reports whether the output was an object or text (workflow_object or workflow_text) and whether it was sent or flagged (send or flagged). A reviewed result is registered through a follow-up call, and a flagged review upgrades the stored report.
Missing handback: when no handback was made, Claude Code records why, such as a refused handback call, no final text, or a plain-text ending.
Why
Workflow agents now return a deliberate final report instead of whatever they said last, and that report can be checked outside auto mode. If a workflow result looks different, it may now come from the handback tool.
