The Action Boundary: Designing Tools, Permissions, and Reversible Effects for AI That Acts on the Real World
Somewhere in your stack there is a line where a model's sampled text stops being text. On one side, a plausible string. On the other, a row changed in a database, money moved out of an account, an email landing in a customer's inbox. Every team building agents has that line. Almost no team has designed it.
The Action Boundary treats that line as an engineering surface with its own discipline, and it delivers three things. Tool design for a stochastic caller: interfaces shaped for a caller that learned your tool by sampling it, not by reading your docs. Authority design: capability envelopes derived from the task at hand, with approval built in as a typed tool call instead of a meeting. And reversible effects: idempotency, previews, undo, compensation, and staged commitment, all carried by a contract that declares what an action does before it does it.
Three running systems, a refunds agent, an infrastructure agent, and a customer-messaging agent, carry every pattern across three consequence profiles, so you see each idea where the stakes are low and again where they are not. By the last page you can walk into Monday's design review and reject the tool that declares no effects, the agent holding standing credentials for a bounded task, and the irreversible write that ships without a preview. It is written for the tool author, the platform owner, and the reviewer who signs off on all of it.
Volume 3 of The AI-Native Builder Canon.
Customer Reviews
ThriftBooks ® and the ThriftBooks ® logo are registered trademarks of Thrift Books Global, LLC
