Skip to main content
The scariest moment in agent building: you have an agent that works, and you need to change it. What if the edit breaks something that took days to get right? Versioning removes that fear. Every change to your agent creates a version — and you can restore any previous one at any time.

​
Two ways to restore

1. Directly in the chat. Every message in the build conversation has a Restore to this version link — click it to bring the agent back to exactly that state.
The agent chat with Restore to this version links under each message

Restore to this version — right where the change happened.

2. Via the Versions tab. Click Workflow in the top right and open the Versions tab — you’ll see the full version history with timestamps and what changed, and can restore any prior version from there.
The Versions tab showing the version history of an agent with restore options

The Versions tab: the agent's full history, restorable at any point.

​
The safe experimentation loop

1

Know your baseline

Your current working state is a version you can return to — that’s what makes the next step risk-free.
2

Make the change

Rewrite instructions, swap tools, restructure the workflow — experiment freely.
3

Test against real examples

Run the changed agent on the same examples that worked before, so you know which version is better.
4

Keep it or roll back

If the new behavior wins, it becomes your new baseline. If not, restore the previous version and try differently.
Once an agent runs reliably, that state is never more than one restore away — which is what lets you keep improving a production agent without treating every edit as a risk.

Watch: Versioning

See how versioning works in the product.

Chapter 3: Scale your craft

Next chapter: sub-agents, skills, and cost control.

Shape agent behavior

What to tighten when a version underperforms.