Re-runs, human approval and public apps
Run one node again without paying for the rest, pause a workflow for a person's approval, publish it as a form, and see what every node costs.
Updated September 26, 2026 · 3 min read
Re-run from a node
Adjusting one agent shouldn't mean paying for the whole workflow again. After a run, right-click the Root or a Sub-Agent on the Canvas (or use the button at the bottom of the inspector) and choose Re-run from this node.
- That node and every Sub-Agent below it call the engine again.
- Everything else is replayed from the last run at no cost: a step whose request is exactly the same as last time (same rules, tools and input) reuses its answer.
- The agents above it re-run only the part that uses the new answer; the part before comes from the last run.
- The console says which steps were replayed and how many tokens that saved.
| Badge on the node | Meaning |
|---|---|
| ● Cached | Unchanged since the last run: a re-run replays its answer. |
| ● Changed | Edited since the last run, or below an edited node: a re-run calls the engine for it. |
Human approval
Drop a Human Approval node between an agent and the Sub-Agent it hands work to: Root or Sub-Agent → Human Approval → Sub-Agent. Its approved output leads to the Sub-Agent it guards; the optional rejected output leads to one that handles a refusal.
The run pauses
When the agent hands the task on, the run waits and a dialog opens with the task and what the reviewer should check (the node's instructions).
The person decides
Approve & continue sends the task on, as written or as edited in the dialog. Stop the run ends it; with a rejected output connected, Reject sends the task there instead.
No answer
After 10 minutes without a decision, the hand-off is rejected.
- Gates also work in Simulation, so the flow can be tried without a key.
- In the exported Claude Code project, the agent is told to show the person the task and wait for their approval before the hand-off.
- A workflow with a Human Approval can't be published as a public app: nobody approves for a visitor.
Public apps
Publish a workflow as a simple page anyone with the link can use: they fill a form, the workflow runs, they read the final answer.
Mark what to ask
Write
{{variable}}in the Root's or a Sub-Agent's behavior rules, for example{{product_name}}. Each variable becomes a form field.Publish
Publish app on the top bar: name the fields (short text, long text or a text file), set the runs per day, and publish. You get the link.
Update or unpublish
The app runs the version saved in your account when you published. Update it after editing; unpublish to kill the link at once.
| Protection | How it works |
|---|---|
| Your Claude key | Public runs are Live runs of your organization: they spend your key and count toward your plan's monthly Live runs. |
| Runs per day | Past the number you set, the app pauses until the next day (UTC). |
| Per visitor | A few runs per minute per visitor, and a cap per app. |
| Connected systems | Off by default: Skills and Connectors are left out of public runs, so a visitor's text can't make an agent act on your systems. |
| What visitors see | The form, a progress count and the final answer; never the workflow, its rules or your key. |
Cost and speed on the Canvas
After a run, each agent shows its time, tokens and estimated cost. Heatmap on the top bar colours the nodes by tokens spent: green under 500, amber up to 4,000, red above.
- The cost is an estimate at the engine's list price, without prompt caching or discounts; engines without a published list price show tokens only.
- The time excludes waiting for a person at an approval.
- The same figures, per run and per node, feed the Usage page.