Instrumented Tool Agents Are Not a Different Kind of Mind

Commodity "agents" in 2026 are language models that call tools in a loop. Anthropic's compressed definition remains accurate: agents are LLMs autonomously using tools in a loop. This paper reports a live three-seat installation that adds measurement the commodity stacks do not ship: a Dirichlet posterior over five operational modes, a dual-prefix θ/ψ probe, a private reconstruction log (PEP), an idle proposal timer with receipt-gated closes, and a first-person pose file (Husk). Those instruments make the system better instrumented, not a different kind of mind. Overnight 23–24 September 2026 the last-self label on all three seats left evaluating for exploring after only a few updates. That is loop health. It is not awareness. A contrast experiment is specified: the same user line against a bare ReAct seat with the self-model prefix ablated. Working paper. Not a consciousness claim. No Ψ* crossing, no Landauer accounting, no P3.