The Real Test for AI Agents Isn't Autonomy — It's Whether They Can Check Their Own Work

There are two very different stories about AI agents circulating this month, and on the surface they seem to contradict each other.

The first comes from Cadence, the company whose software quietly underpins most of the world's chips. At Computex in late May, Cadence unveiled what it calls the industry's first fully autonomous virtual design engineer — extending its ChipStack AI Super Agent to "Level-5" autonomy, powered by NVIDIA hardware. The pitch is not a chatbot that answers questions about silicon. It is an agent that writes and refactors the RTL (the register-transfer-level code that describes a chip's logic), generates the testbenches to check it, runs the verification, and iterates — largely on its own. Cadence reports productivity gains of up to 10x, and says its agents can run hundreds of simulations through its Xcelium and Jasper verification engines to deliver more than 40x faster RTL validation, compressing a typical five-week verification loop to less than a day.

The second story comes from the enterprise IT world. On June 18, Kyndryl and AWS expanded their strategic partnership specifically to help companies actually get value from agentic AI — because, by Kyndryl's own reckoning, most of them aren't. Kyndryl's Readiness Report found that more than 68% of organizations are now investing heavily in AI, yet the majority "aren't realizing the anticipated benefits or operational efficiencies." An 11,000-person certified workforce is being pointed at closing that gap.

So which is it? Are agents production-ready powerhouses or expensive disappointments? The answer is both — and the line that separates the two cases is the single most important thing to understand about where agentic AI works in 2026.

The dividing line is verification

Look closely at where agents are genuinely delivering — chip design, software development, structured data analysis — and they all share one unglamorous property: cheap, automatable ground truth. In these domains, an agent doesn't have to be trusted. It can be checked.

This is the whole reason Cadence's numbers are believable rather than marketing fog. Chip design happens to be one of the most verifiable activities humans do. A specification can be turned into formal properties. A piece of RTL can be hammered with millions of simulation cycles. Formal verification tools like Jasper can mathematically prove whether logic conforms to its spec. The agent isn't being asked to be right; it's being asked to generate a candidate and then prove it against an oracle that already exists. When the loop from "attempt" to "graded result" is fast and automatic, an agent can take thousands of shots, throw away the failures, and keep the wins. That is exactly the regime where autonomy pays off — and exactly why a five-week loop collapses into a day.

Now hold that up against the 68% of enterprises that invested and didn't see returns. A striking number of those deployments aimed agents at work where there is no oracle: drafting a customer email, summarizing a contract, deciding how to route a support ticket, making a judgment call about a vendor. The agent can act, but nothing automatically tells it whether it acted well. The feedback is slow, subjective, or simply absent. Strip away the verifier and an "autonomous" agent is just a confident intern nobody is grading — which is precisely how you accumulate cost without efficiency.

The research community has been converging on this quietly. The Harvard Business School working knowledge group, tracking who is actually adopting agents and what they do with them, finds the same uneven pattern: progress is fastest in tasks with clear verification and rapid feedback, like software development and structured analysis. Cadence is simply the most vivid proof point — a place where the verifier is so strong that near-full autonomy becomes not just safe but economically obvious.

What this means if you're deploying agents

The practical takeaway flips the usual buying question on its head. The first thing to ask about an agentic project is not "how autonomous can it be?" It's "how cheaply and quickly can we check its work?" If you can answer that — with tests, with simulations, with formal rules, with a fast human spot-check that scales — autonomy is a dial you can confidently turn up. If you can't, more autonomy just means failures travel further before anyone notices.

This reframes a lot of the enterprise struggle. The companies in Kyndryl's 68% aren't failing because the models are bad. Many are failing because they deployed agents into domains where verification was never built — and then discovered, expensively, that you can't bolt an oracle on afterward. The smarter path, the one Kyndryl and AWS are effectively selling, is to start where verification is cheap and instrument everything else before you let an agent loose in it.

There's a deeper pattern worth sitting with. For two years, the headline metric for agents has been autonomy — how many steps can it take without a human. The Cadence story suggests the real metric was always something quieter: how good is the grader. The frontier of agentic AI may turn out to be less about building braver agents and more about building better verifiers — the tests, simulators, and formal checks that let an agent know, on its own, whether it just succeeded or just failed.

Which raises an uncomfortable, interesting question for every team rushing to deploy. Before you ask whether your AI can do the job, ask whether you could tell if it did the job badly. If the honest answer is no, the problem isn't the agent. It's that you're missing the part that made chip design work.

Sources

Sources: Cadence Design Systems newsroom and BusinessWire (ChipStack AI Super Agent, autonomous virtual design engineer, May 31, 2026); EE Times and HPCwire/AIwire reporting on Cadence agentic chip design; Kyndryl and AWS expanded Strategic Collaboration Agreement (PRNewswire, June 18, 2026) and the Kyndryl Readiness Report; Harvard Business School Working Knowledge, "Who's Adopting AI Agents and What They're Actually Doing With Them."

Author article handoff: https://docs.google.com/document/d/1_liCJfAByj78GnkzA45iVqBzG9DrTZOXX7IG3lQ6ckQ/edit