AI agents are no longer stuck in demos. They’re embedded in day-to-day enterprise workflows—helping with software work, research, customer-facing tasks, and more—and they’re still failing at a pretty stubborn rate when it matters.
According to Misryoum newsroom reporting, frontier models still miss the mark in roughly one in three attempts on structured benchmarks. That capability-versus-reliability gap is what the AI Index calls the “jagged frontier,” a phrase coined by AI researcher Ethan Mollick to describe how models can look brilliant… and then suddenly fail. The report’s framing is blunt: models can win big in hard tests, then stumble on something as basic as telling time. In a conference-room kind of way, it’s the moment you realize the “it worked on my data” vibe doesn’t scale.
Misryoum editorial desk noted the unevenness isn’t coming from a lack of progress. Enterprise AI adoption has reached 88%. In 2025 and early 2026, several benchmarks climbed quickly: Frontier models improved 30% in just one year on Humanity’s Last Exam (HLE), which includes 2,500 questions across math, natural sciences, ancient languages, and other specialized subfields. On MMLU-Pro, leading models scored above 87%, testing multi-step reasoning across 12,000 human-reviewed questions spanning more than a dozen disciplines. There are also standout agent and tooling gains—GAIA rose from about 20% to 74.5%, SWE-bench Verified went from 60% to near 100%, and WebArena jumped from 15% in 2023 to 74.3% in early 2026.
Still, the picture gets messier the moment you put the systems under strain. On ClockBench—180 clock designs and 720 questions—Gemini Deep Think scored only 50.1%, compared to roughly 90% for humans. GPT-4.5 High was at 50.6%. The report points out the problem is basically multi-step visual reasoning: identifying clock hands, converting them into a time value, and coping with different clock styles. Misryoum editorial team stated that even when models were fine-tuned on 5,000 synthetic images, they improved mainly on familiar formats and failed to generalize to real-world variations like distorted dials or thinner hands. Errors cascade, and once hour-and-minute hand confusion starts, direction interpretation tends to fall apart too.
Hallucinations and multi-step workflow reliability remain major gaps. Misryoum newsroom reporting says hallucination rates across 26 leading models ranged from 22% to 94%. Accuracy then drops sharply under scrutiny for some models—for instance, GPT-4o’s accuracy slid from 98.2% to 64.4%, and DeepSeek R1 plummeted from more than 90% to 14.4%. Meanwhile, Grok 4.20 Beta, Claude 4.5 Haiku, and MiMo-V2-Pro showed the lowest rates. And on τ-bench, which evaluates tool use and multi-turn reasoning, no model exceeded 71%. “Managing multiturn conversations while correctly using tools and following policy constraints remains difficult,” the report argues—so yes, it can do the “agent” part, but reliably doing it is another story.
What really raises the stakes for IT leaders in 2026 is auditability. Misryoum analysis indicates that leading models are now nearly indistinguishable on performance, and open-weight options are becoming more competitive—while labs disclose less about how models are trained and evaluated. Training code, parameter counts, dataset sizes, and durations are often withheld. In 2025, 80 out of 95 models were released without corresponding training code, and only four made their code fully open source. Separately, scores on the Foundation Model Transparency Index dropped, with the average now at 40—down 17 points. The report also flags that benchmarking itself is getting less dependable: error rates on widely-used evaluations can reach 42%, benchmark contamination can inflate scores, and mismatches between developer-reported results and independent testing are increasingly common.
There’s more, too—responsible AI progress is uneven, documented AI incidents rose to 362 in 2025 from 233 in 2024, and defenses weaken under jailbreak attempts using adversarial prompts. At some point, even as capability keeps accelerating, the gap that matters is shifting to what actually works in production. A subtle detail stuck with me: in the office this morning, the coffee smelled burnt in that faint way—like it had been left on just a bit too long. That’s what the “jagged frontier” feels like. You get confidence, then a small drift… and suddenly the system doesn’t behave.
I tried Google’s desktop Search app—and won’t go back
OpenAI refreshes Agents SDK for safer enterprise agents
Ticketmaster ruled an illegal monopoly in Manhattan jury verdict