#agent-architecture
3 items
What Is mupot? The Operating Layer You Keep Re-Deriving
Ask any frontier model to design an AI team that runs a business website and it returns the same seven things — context, memory, tools, validation, approvals, tasks, agents. That convergence is the tell. It's not a framework to build. It already exists.
From Approval to Action: The Executor That Re-Binds to the Proposal
Most agent frameworks let the act step trust what the caller sends at execute-time, not what a human approved. That's prompt-injection-at-execute. The fix is one rule: the executor re-binds to the stored proposal — the caller passes only an ID.
Shipping a Brain to Someone Else's Box
We tried to install a sovereign AI brain on a customer's own server. Every hidden assumption of running it on our own machine came due at once. A field report — the setbacks, the drama, and the moment it finally thought with the customer's own cloud.