Robert Ta finds 591 failures

AI agent failures are starting to look less like “the model said something dumb” and more like ordinary software disasters with a language model in the middle. That(clyro.dev)cidents from 2023 to 2026. The claim is blunt: 88% of classifiable failures trace back to infrastructure gaps rather than pure model quality. In other word(clyro.dev)t is. (clyro.dev) ### What kind of failure is this? Clyro draws a clean line between LL(clyro.dev)gent failure is a bad action chain — the system reads stale context, picks the wrong tool, keeps retrying, mutates memory, or quietly drifts while every dashboard still says green. That matters because the blast radius is completely different. One bad answer is annoying. One bad action loop can run for days. (clyro.dev) ### What were the biggest buckets? The dataset groups failures into five (clyro.dev)t 30.3%. Silent Degradation accounts for 24.9%. Then come Memory Corruption at 8.1% and Runaway Execution at 5.1%. The ordering is the interesting part — the most common failures are not spectacular robot meltdowns. They are quieter operational misses where the agent lacks the right state, takes an action it should not, or gets worse without tripping alarms. (clyro.dev) ### Why is context(clyro.dev)an old return policy, misses a recent handoff note, or loses the constraint that mattered two steps ago, the next action can still look internally reasonable. But it is wrong in the real world. Basically, the model is driving with a dirty windshield. Better reasoning does not help much if the agent cannot see the current situation. (clyro.dev) ### What counts as a rogue action? This is when the agent does something it technically can do but s(clyro.dev)instead of reads, tool calls without approval, or business-logic violations that pass because the system only validated syntax. That is why this category sits so high. Permission design, policy checks, and scoped tools are boring infrastructure work — but turns out they are exactly where a lot of the damage starts. (clyro.dev) ### Why does silent degradation mat(clyro.dev)t fail loudly. They degrade while returning HTTP 200, while traces look normal, and while nobody gets paged. The article’s example of a retry spiral that kept burning money for 11 days before anyone noticed captures the point. A system can be “up” and still be failing in the way that matters — cost, quality, and downstream actions. (clyro.dev) ### So is the model off the hook? No — but it is not the mai(clyro.dev)tations, while 22.5% are mixed or unclear. That means model quality still matters, but the bigger operational win is elsewhere. If your agent keeps making bad moves in production, swapping in a smarter model may help less than adding session isolation, step limits, cost ceilings, runtime policy enforcement, and human review on risky actions. (clyro.dev) ##(clyro.dev)ligence. Clyro’s own write-up points to rising failure and abandonment rates across AI projects more broadly, and uses this dataset to argue that teams are measuring the wrong bottleneck. The hard part is not only getting an agent to think. It is getting the agent to operate safely, observably, and with bounded authority in production. (clyro.dev) ### Bottom line? The practical takeaway is simple. Treat agents less like chatbots and mo(clyro.dev), no audit trail, and infinite retries. A lot of agent stacks still do exactly that. (clyro.dev)