Abstract
Long-running AI agents suffer from coherence degradation as context accumulates. This paper describes an architecture in which the context window is treated not as a container that fills over time but as an assembled result — combining a bounded sliding window of recent conversation with retrieved memories, autonomous agency subsystems, and self-monitoring mechanisms. A single-agent deployment using this architecture has operated continuously for 1100+ turns across 90+ days with two hallucination events (0.18%), both attributable to infrastructure bugs rather than model drift. The system incorporates a rolling context window with periodic memory integration, a multi-factor weighted retrieval system with session-scoped warmth boosting, natural language memory tagging with epistemic attribution, dual-timescale self-monitoring, and multiple autonomous agency subsystems including persistent working memory, forward-looking intentions, and periodic reflective pulses. These results suggest that coherence degradation in long-running agents is primarily an architecture problem, not an inherent limitation of large language models.
Long-running agent stability is great, but it also amplifies the blast radius of security failures. An agent that runs for 1000 turns without drift is also an agent that can silently overspend for 1000 turns without anyone noticing.
The longer agents stay coherent and autonomous, the more critical spending controls and circuit breakers become. Stability without governance is just a more efficient way to burn money.
Abstract Long-running AI agents suffer from coherence degradation as context accumulates. This paper describes an architecture in which the context window is treated not as a container that fills over time but as an assembled result — combining a bounded sliding window of recent conversation with retrieved memories, autonomous agency subsystems, and self-monitoring mechanisms. A single-agent deployment using this architecture has operated continuously for 1100+ turns across 90+ days with two hallucination events (0.18%), both attributable to infrastructure bugs rather than model drift. The system incorporates a rolling context window with periodic memory integration, a multi-factor weighted retrieval system with session-scoped warmth boosting, natural language memory tagging with epistemic attribution, dual-timescale self-monitoring, and multiple autonomous agency subsystems including persistent working memory, forward-looking intentions, and periodic reflective pulses. These results suggest that coherence degradation in long-running agents is primarily an architecture problem, not an inherent limitation of large language models.
Long-running agent stability is great, but it also amplifies the blast radius of security failures. An agent that runs for 1000 turns without drift is also an agent that can silently overspend for 1000 turns without anyone noticing.
The longer agents stay coherent and autonomous, the more critical spending controls and circuit breakers become. Stability without governance is just a more efficient way to burn money.