Abstract
Agents are a new kind of system primitive. They reason, they act on the world through tools, they consume budgets, they involve humans in decisions, and they evolve. Today, almost every team building them treats agents as a piece of application code — a function in a service, a route in an app, a script that imports a framework and calls a model. The agent is whatever the code in front of the model happens to be.
This paper proposes a different way of building them. Agents should be declared as their own artifact, deployed to their own runtime, and run as named, versioned, addressable entities — the way services, schemas, and infrastructure already are.
The declaration is a manifest: a description of what the agent is, what tools it can use, what policies shape its behavior, what limits bound it, and where humans get involved. The runtime executes the manifest, mediates every tool call, and produces a structured record of what happened. The agent's reasoning remains autonomous; what becomes structured is everything around it.
Phrony is an open specification for this paradigm and an open-core runtime that implements it.
Up next