Guardrails you write once and attach anywhere
Rules used to live inside each agent's prompt, which meant editing five prompts to change one policy. Guardrails are now first-class objects you author once and attach to any agent in the workflow.
Every team we onboarded hit the same wall in week two. They wrote a careful prompt for their first agent, copied it into the second, tweaked it slightly for the third, and then discovered that changing one policy meant editing every prompt by hand. Rules that mattered lived in prose, scattered across a canvas.
Rules are objects now
A guardrail is a named rule with its own editor, its own version history, and its own attach list. You write Only Answer From Knowledge once, then attach it to the three agents that need it. Change the wording later and all three pick it up on the next run. Nothing gets copied.
- Only Answer From Knowledge keeps the agent inside the documents you approved, and makes it say so when the answer is not there.
- Stay On Topic declines politely and steers back to what the workflow is for.
- No PII Collection refuses to record identifiers the workflow has no business holding.
You can also write your own. A guardrail is just a rule with a name, so the ones we ship are not special.
Attaching from the canvas
Open any agent node and the guardrail panel lists everything in the workspace with a checkbox next to it. The attach list is stored on the guardrail, not the agent, so you can also open a guardrail and see, in one place, every agent currently bound by it. That view turned out to be the one people actually wanted during audits.
The question stopped being "what did we tell this agent" and became "which rules apply here", and that one is answerable.
What this changed in practice
Prompts got shorter. The median agent prompt in new workspaces dropped by roughly a third, because the defensive paragraphs moved out of the prompt and into rules that are visible on their own.
It also made review possible. A compliance reviewer who does not want to read prompt engineering can read four named rules and their attach lists, and get an accurate picture of what the workflow will and will not do.
Migrating an existing workspace
Nothing breaks. Existing prompts keep working exactly as written. When you open an agent, the builder highlights sentences that look like policy and offers to extract them into a guardrail, and you decide one at a time whether that is what the sentence was for.