ML//agent//agent harness//epistemic interface

An agent's tools are its epistemic interface: they define the observations from which its world is constructed. A programmer already sees a repository only through an editor, a terminal, search results and test output (never by loading every physical bit into consciousness), and an agent is even more constrained.

An agent's tools are its epistemic interface: they define the observations from which its world is constructed. A programmer already sees a repository only through an editor, a terminal, search results and test output (never by loading every physical bit into consciousness), and an agent is even more constrained.

ot=OH(st)o_t=O_H(s_t)ot​=OH​(st​)

The environment state sts_tst​ may contain files, processes, Git history, and network services. The harness observation function OHO_HOH​ turns some of that state into serialized context oto_tot​, and the model's policy chooses its actions from those partial observations, never from sts_tst​ itself. Existing data is not known data.

Read, search, and inspect tools behave like sensors. Write, delete, deploy, and send tools behave like actuators. Tests do both: they alter computational state and return evidence.

This is why tools are epistemic interfaces, not merely capabilities. They determine what distinctions the model can perceive before it acts.

The phrase is a lens over the observation side of an Agent-Computer Interface, not a claim to a new interface category. Its value is the question it foregrounds: what can this representation let the agent know, and what distinctions has the harness erased before inference begins?