Blog
Notes on putting agents on rails
How we think about governing autonomous systems: the rail model, the method behind it, and what we learn running it.
-
Announcement
6 min
Introducing Airbender ARC
Agents will be wrong some of the time. Airbender ARC is the controller that decides what happens next, enforces it in the request path, and keeps the evidence.
-
Rail model
7 min
Rails, not guardrails: deciding what happens after an agent is wrong
Filtering text catches some failures. Governing actions catches the expensive ones. How ARC turns many rail verdicts into one decision, and what it does when it can't decide at all.
-
Method
8 min
HAZOP for agents: nine guide words that find what evals miss
The weakest point of a governance platform is the rails nobody thought to write. Process safety solved that problem decades ago. Here is how ARC borrows the method.
-
Evidence
6 min
Evidence you can replay
An audit log tells you what happened. A replayable verdict lets an auditor run the decision again and get the same answer. Why ARC stores the second kind.
-
Architecture
6 min
No activation, no AI: the contract of operations
Governance that isn't established at connection time erodes at runtime. How an app proves to ARC that it implements what the released rails demand before it gets a credential.