The article introduces the Supervisor Tax problem in multi-agent AI systems, highlighting context leakage from free-form outputs, evaluation drift with potential infinite loops, and the lack of auditable receipts.

Illustrative implementation shows an XState-based pipeline with defined states—plan, execute, verify, repair, finalize, escalate_human—and guards that enforce deterministic transitions.

Deterministic guardrails encode constraints as code-level transition guards, enabling zero-token supervision between steps and hard termination when conditions are not met.

The approach delivers zero supervision prompt tokens, guaranteed loop termination via guards, and isolated working contexts for each step to reduce noise and token usage.

The traditional supervisor LLM is replaced with deterministic typed state machines that route and govern agent steps, cutting token waste and increasing determinism.

Receipts become immutable, validated outputs that capture essential execution details and deterministic transitions, enabling verifiable provenance.

In production across hundreds of tasks, token use dropped dramatically, median latency fell from the high forties to the mid-teens seconds, infinite-loop faults vanished, and auditability shifted to SQL/JSON queries over state transitions.

Agents run state machines while routing is handled by the machine; each leaf agent returns a strongly typed, schema-validated receipt with status, duration, tokens, nextTrigger, and artifacts.

Takeaway: Reserve LLMs for creative and complex reasoning, while control plane tasks like routing and coordination should rely on deterministic state machines and verifiable receipts.

Call to action: Readers are invited to share experiences or questions about preventing agent routing drift.