Super Agent, the execution debugger, and human approvals

Workflows can now plan their own work, pause for a person, and be inspected while they run.

We're introducing the Super Agent node: a single node that breaks a task down and orchestrates its own sub-agents at runtime, instead of following a sequence you laid out by hand. It's live in the production app today.

Alongside it:

• Execution debugger: watch a run live and inspect variables, the blackboard, and agent memory at every step
• Human approvals: add an approval checkpoint to a workflow, or let an agent request one mid-run as a tool
• Documents, end to end: agents read charts and diagrams inside PDF and Office files, and scans and slide decks now extract properly

Agent memory recall is now advisory and screens out unverified facts, so agents treat what they remember as a hint rather than the truth.

We also closed server-side request forgery risks in SSO discovery, webhook delivery, and image fetching, and access tokens are invalidated the moment you log out.