Delay-Induced Instability in Multi-Agent LLM Orchestration: Exact Stability Boundaries and a Delay-Aware Extension of UCCF

Orchestrated systems of LLM agents read a shared state, compute a correction and write it back, and every step of that loop takes time. This paper extends the author’s earlier UCCF preprint with delay. For a linear error-correction model we give exact stability boundaries in three senses (mean, mean-square, almost sure) and show that they differ by large factors under random delays, sign-flip errors and persistent latency regimes. We prove that the worst-case delay margin is a sufficient and tight rule for heterogeneous agent delays, show that the margin survives saturation but not a second-order plant, and show that when delay grows with load the safe gain peaks at an intermediate team size. We extend the published UCCF toy plant with actuation and observation delay and N agents, and present a delay-aware controller (UCCF-DR). Of 26 pre-registered claims, 24 hold and two fail narrowly; both are reported. Everything is simulation except one exploratory latency measurement of a real orchestration backend (OpenAI gpt-4o-mini), which showed multi-second delays, a weak load effect and no persistent slowdown regime in runs of a few minutes. Code, data, claim hashes and the latency traces (timestamps only) are included.