A scientific kernel for agentic architectures

Agent runtimes and agent operating systems know who may do what. A kernel for scientific work must know three more things: what kind of number it is looking at (exact, estimated or generated), whether a claim reproduces before it enters the world model, and when the world is about to change. We report a kernel built around these three questions. Agent operating systems treat each agent as an application scheduled by a model-centred kernel. We invert that assignment for regulated, traceable settings: the agent loop is implemented once, as the kernel; the model is a resource it calls; an application contributes a workflow graph, its gate policies and a window, and no substrate machinery. The kernel owns four properties: durable runs that pause at a gate for days and survive restarts; principals with grants enforced per tool call; gates as run states; and one provenance activity per tool call, carrying the epistemic class of the result. Deterministic tools may abstain; abstentions are first-class objects. Several agent architectures - a bounded chat, a discovery campaign, a design-make-test-analyze loop - run as graphs on one engine and appear to the scientist as one list of waiting decisions. Nineteen of the twenty-three executable tests are measured in an automated nightly run against the frozen executable, among them durability across restarts, a two-hour wait holding no thread, a foreign agent held at a gate, and a new architecture added with zero kernel changes; the four that are not measured in the cited run are named with their cause.