HUAP: Toward a Trace-First Agentic Operating System A public core architecture for reproducible agent workflows

The current LLM agents ecosystem is rich in prompting patterns and tool-use wrappers, yet weak in operational guarantees: reproducibility, regression testing, auditable execution, and principled resource accounting. This paper proposes an agentic operating system viewpoint and presents HUAP (HUman Agentic Platform) as a trace-first public core architecture. HUAP treats agent workflows as executable graphs, tool calls as OS-like syscalls, and execution traces as the primary artifact for debugging, evaluation, and certification. We describe the kernel services needed for real-world agentic systems-pod contracts, deterministic replay, trace diffing, and budgeted evaluation-and show how these services turn "agent behavior" into a testable software system. We relate the design to durable execution and event sourcing, and position HUAP relative to contemporary agent frameworks and tool-use training paradigms.