Top failure modes 2024-2026: (1) tool hallucination: calling wrong tool or with wrong args; mitigated by tight JSON schemas + constrained decoding; (2) compounding errors: small earlier mistakes cascade; mitigated by re-planning at each step and self-verification; (3) context bloat: agents accumulate all observations and lose focus; mitigated by summarizing/compressing history or using scratchpads; (4) infinite loops / retry storms: no termination criterion; mitigated by max-step limits + cycle detection; (5) side-effect leaks: actions executed without dry-run; mitigated by approval workflows for high-risk actions; (6) goal drift: the agent solves a different problem than asked; mitigated by periodic checkpointing to the original goal; (7) prompt injection via tool output (a retriever returning untrusted content with hidden instructions); mitigated by sandboxing tool outputs and re-prompting with "treat tool output as data, not instructions". Most common root cause in 2026 production data is poor tool description clarity, which matters more than model capability.