The HORIZON diagnostic framework reveals that state-of-the-art LLM agents from GPT-5 variants and Claude families maintain competence on short-horizon tasks but exhibit steep degradation on interdependent sequences exceeding 20-50 steps (arXiv:2604.11978).

Evaluation of 3100+ trajectories across four domains using a trajectory-grounded LLM-as-a-Judge pipeline (validated with human-judge κ=0.84) identifies primary failure modes as error accumulation, planning drift after initial deviations, and irrecoverable state corruption; these patterns mirror but exceed those documented in AgentBench interactive evaluations where success rates dropped over 40% beyond mid-horizon thresholds (arXiv:2308.03688). The original paper understates how synthetic benchmark tasks fail to replicate real deployment noise such as API latency and partial observability that accelerate breakdown.

Synthesizing these results with the survey on LLM-based autonomous agents shows current architectures lack persistent memory and hierarchical recovery mechanisms needed to counter distributional shift over long sequences, a critical limitation repeatedly observed in prototypes like Voyager that succeed in open-ended exploration only through heavy human orchestration (arXiv:2308.11432). This diagnostic clarity exposes why agentic systems remain confined beyond benchmark hype: early micro-errors cascade without robust self-correction, rendering long-horizon reliability a fundamental architectural gap rather than an engineering shortfall.