What Breaks in Long-Lived Agents: An Eight-Month Operating Record of a Ten-Agent Fleet

Most agent benchmarks score a task in a fresh context, and newer memory benchmarks probe recall across sessions. Some costly failures can occur in one session but become visible and accumulate only over months. We report eight months of operating a fleet that grew to ten long-lived coding agents (nine on Claude Code) on four machines. First, false zeros: empty tool results taken as proof of absence. We list eighteen carriers and quantified cases such as 64 of 1,501 memory files silently dropped at a permission boundary. Second, verification needs a different position or instrument: a guard change that passed its author's check let through 52 commands its predecessor blocked. Third, we measure where a 1M-token window goes. The model's own output is the largest share (39%), and tool-server responses are below 1%. After a client update, idle compaction began to fire, visible only in its timing. A pre-registered blind study shows a full handoff brief restores 70.7% of probed facts after rotation, against 0% without it, but loses to the native summary under the same character limit. Fourth, we report negative results, including a mood classifier that did not beat "always neutral". Preprint, not peer-reviewed. The Study C reproducibility package (per-unit counts, frozen scorer, rebuild script and tests) is at https://github.com/rushnur88/what-breaks-in-long-lived-agents.