The Limits of Prompt-Based Self-Improvement in Language Model Agents: Meta-Optimization and Failure Attribution

Abstract: Modern language models' in-context learning and instruction-following abilities make it possible to adapt agent behavior without updating model weights. This thesis studies online self-improvement, where an agent encounters each task only once and must accumulate reusable experience across a sequence of tasks, reflecting real-world deployments in which actions have persistent consequences and retries may be impossible. First, we ask whether meta-optimization can be integrated with online self-improvement by allowing the prompt optimizer itself to learn across episodes. Building on Agentic Context Engineering (ACE), we evaluate three variants that equip its optimizer with persistent memory. On the AppWorld benchmark, none reliably improves over ACE when using strong base language models. Analysis suggests that ACE-style adaptation is near saturation in this regime, leaving little room for meta-optimization to provide further gains. To probe the limits of ACE-style adaptation, we evaluate it on Toolathlon, a harder and more heterogeneous tool-use benchmark, and observe no improvement over non-adapting agents. Auditing these apparent failures reveals ambiguous task specifications, missing information, brittle evaluators, and unstable external services, rather than cleanly attributable agent errors. Because agent trajectories can span hundreds of thousands of tokens, we develop an automated, evidence-grounded pipeline for analyzing trajectories and attributing failures at scale. Together, these results show that effective online self-improvement depends not only on prompt optimization, but also on transferable structure across tasks and trustworthy failure attribution. Committee: Aran Nayebi (advisor) Daniel Fried Haochen Zhang