RO: Reliable Orchestration of LLM Programs via Deterministic-Generative Separation

Recent failure analyses of multi-agent LLM systems reveal that the dominant failure mode is not reasoning but orchestration: plan abandonment, state corruption, and tool- management errors dominate system-design failures across seven frameworks and 1,600+ execution traces. We trace this to an architectural conflation—current frameworks require the LLM to simultaneously manage deterministic bookkeeping (loops, state, control flow) and non-deterministic reasoning (classification, generation), despite these being fundamentally different computational concerns. We present RO (Reliable Orchestration), an execution engine enforcing strict separation: a typed runtime sequences deterministic instructions while the LLM is invoked only at explicitGOALboundaries requiring semantic judgment. Schema validation at each boundary ensures outputs conform to declared types before entering memory, yielding a central invariant: orchestration correctness is independent of LLM behavior. Programs are authored in natural language and compiled to an auditable DSL. We demonstrate RO by reimplementing a 12,518-line Rust diagnostic agent as a 444-line RO program—a 28×reduction—with equivalent outcomes and2–3×wall-clock speedup on ITBench SRE scenarios.