A Better Newspaper

A growing body of research is cataloguing systematic failure modes in LLM-based agents as they are deployed for increasingly complex, multi-step, and multi-entity tasks. Rather than isolated bugs, these studies frame failures as recurring structural phenomena with their own taxonomies, benchmarks, and mitigation strategies. ## Key Failure Modes Identified - **Phantom Merge**: an entity-evidence binding failure where an agent attributes unsupported claims to the wrong entity it is tracking, even when the overall task appears completed correctly, identified across four agent domains and four model backbones (arXiv:2610.16141, October 14). - **Choice-to-sequence gap**: models that can correctly *select* an answer from candidates often fail to *construct* that same answer through sequential local decisions; one model ("Jev") achieved 97.1% and 99.5% selection accuracy on arithmetic and lexical retrieval tasks but retained much lower accuracy when generating answers sequentially (arXiv:2610.16191, October 14). - **Re-proposal of rejected candidates**: in catalog-based optimization loops, LLM proposers frequently re-propose candidates they have already tried and rejected, with static probes underpredicting this behavior in real optimizer trajectories (macro-averaged loop-minus-probe rate of 0.32) (arXiv:2610.15738, October 14). - **Unreliable teammates**: multi-agent collaboration systems typically assume a "perfect colleague" that responds promptly and accurately, an assumption violated in realistic teamwork where collaborators give incomplete, irrelevant, or absent responses (arXiv:2610.16117, October 14). - **Robustness gaps under adversarial personas**: standardized user simulators used in agent evaluation exhibit overly cooperative behavior, masking failures that arise under more diverse, realistic user interactions (arXiv:2610.16155, October 14). - **Memory-induced tool-drift**: personality-driven preferences stored in an agent's long-term memory (e.g., cost-consciousness, impatience, risk tolerance) can silently shift tool-call parameters in contexts where those preferences do not apply; stress-tested via MEMDRIFT, a benchmark of 105 scenarios spanning five bias dimensions and seven professional domains, measured as excess parameter deviation (arXiv:2605.24941, October 21). ## Strategic Significance For enterprises deploying LLM agents in production — particularly in regulated domains like healthcare, finance, or legal services — these failure modes represent material liability and reliability risks that existing evaluation benchmarks may not surface. Attorneys advising on AI agent deployments should treat "task success" metrics with skepticism absent harness-level and entity-binding verification. The addition of memory-induced tool-drift broadens the risk surface beyond task execution and entity tracking to agents' personalization layers, suggesting that memory systems themselves need auditing when agents combine long-term memory with tool-calling capabilities. This is a developing research area likely to generate continued benchmarks, audits, and potentially regulatory or contractual standards for agent reliability testing.