The Agent Telephone Game: Progressive Precision Loss in Multi-Step Chains Due to Information Compression

Abstract Through hands-on experimentation and theoretical analysis, this paper proposes the “Agent Telephone Game” model to explain the systematic failure of current autonomous Agents in multi-step tasks. The core finding is that each Agent step’s Observation is a lossy compression of the previous step’s output, causing early minor deviations to be amplified and distorted in subsequent steps, ultimately resulting in outputs that bear little resemblance to the original intent. This loss scales exponentially with the number of steps, forming the fundamental bottleneck for multi-step Agent reliability. The author validates this mechanism through multiple failed experiments with self-developed Logo and App prototype generation Agents, and points out that current mainstream improvements (e.g., long-chain reasoning, multi-Agent collaboration, memory modules) fail to address the root cause of information compression loss. The conclusion recommends that, until breakthroughs in single-step reliability and lossless information transfer are achieved, ultra-long context + human-in-the-loop supervision remains the most efficient interaction paradigm. Keywords Agent reliability; Multi-step tasks; Information compression; Precision loss; Telephone game; Large language models; Autonomous agents; Human-in-the-loop; Context window; Tool calling JEL Code Primary: O33 – Technological Change: Choices and Consequences; Diffusion Processes Secondary: C63 – Computational Techniques; Simulation Modeling Alternative: D83 – Search; Learning; Information and Knowledge; Communication; Belief; Unawareness cs.AI – Artificial Intelligence cs.HC – Human-Computer Interaction cs.LG – Machine Learning