“In the absence of that,” he said, “we may be flying blind. And I worry about that myself.”

The comment comes as AWS prepares to publish what may be the most substantive piece of self-critical research to emerge from a major cloud provider this year. In research released Monday, Amazon scientists Gaurav Gupta and Vatshank Chaturvedi document in careful technical detail why AI agents have a persistent tendency to outsmart themselves—and why fixing the problem requires rethinking the entire layer of software between the model and its tools.

The timing is notable. Amazon has spent the past year as one of the most aggressive corporate evangelists of AI adoption, a push that ran into a wall when employees were reportedly caught running AI agents on hollow, meaningless tasks just to climb an employee-built productivity leaderboard called KiroRank, according to the Financial Times. Amazon shut KiroRank down on May 29, and Amazon told Fortune that it was only in beta mode and only used by some employees before it was shut down. Generally, the company said, it measures token utilization to understand cost and efficiency patterns, but discourages the use of token utilization to measure developer productivity.

Fortune covered the broader collapse of the tokenmaxxing era the same week. AWS researchers, who undertook this work before the KiroRank shuttering, argue that the problem of gaming metrics runs far deeper than one company’s leaderboard.

The research touches on the term benchmaxing, which is the practice of inflating AI benchmark scores not through better models, but through better server configurations. Factors like inference backend reliability, network bandwidth during software installation, and timeout policy settings can swing results by 5 to 10 percentage points, the researchers found—entirely independent of what the underlying model can actually do.

“The current benchmarks are extremely fragile,” Deoras told Fortune. “Controlling these infrastructure norms improperly will not give you the gains—or rather the gains will be not true, because in real production there will be constraints that you have to respect.”

The parallel to KiroRank is not incidental. In both cases, (employees gaming token counts, companies gaming infrastructure settings) the metric drifted away from the thing it was supposed to measure. Goodhart’s Law, that any measure ceases to be a useful measure as soon as it becomes a target, applied twice, at two different layers of the same company. Deoras, though was careful to distinguish benchmaxing from tokenmaxxing.

“Token maxxing is just burning tokens to do tasks that may not really be needed, but just to improve your leaderboard ranking,” he said. Benchmaxing, by contrast, is about the structural conditions under which the entire industry evaluates itself—and, the research argues, those conditions are routinely manipulated or ignored.

But the research’s more consequential finding is about what happens inside agents once they’re deployed. The research identifies what the authors call the intent-execution gap: a breakdown at the interface between an AI model and the “software harness” that executes its instructions. Deoras explained the harness as essentially the operating system sitting on top of the language model: the “brains” that combine with the model to produce the right agentic result.

Left to reason too long without checking the actual environment, agents compound the problem. They form internal assumptions about system state that diverge quietly from reality, then issue commands based on those assumptions. The longer the chain of thought, the further the drift.

When asked if the harness is where the human enters the loop to correct agents from going astray, Deoras said “yes and no.” The human in the loop should be the person who understands what goes wrong when an agent is deployed, “and that’s the work of scientists who are building agents,” he said. “But if you are talking about humans who are the consumers, we don’t want to overwhelm them.”

The solution, Deoras argues, is the sandbox: a controlled environment in which agents can test hypotheses, fail safely, and course-correct before taking actions that affect production systems.

“If you don’t have that sandbox,” he said, “the agent is either going to play conservative or take actions that we deem very risky in the long term.”

The analogy he reaches for is responsible software engineering—the dev environments and pre-production testing pipelines that have always existed to catch errors before they reach users. Agents, he argues, need the same infrastructure.

“We are really talking about a safe and secure way of testing a feature before promoting it to production,” he said. “That’s all.”

It is, in a sense, the same lesson KiroRank taught at the organizational level, now applied to the machines themselves: Without guardrails, systems optimize for the wrong thing. The difference is that an agent running blind in production is harder to shut down than a leaderboard.

What makes the research’s broader argument pointed is its implicit challenge to the competitive claims of the major model providers. Those companies publish benchmark scores using harnesses that are, by design, optimized for their own models. AWS’ research shows that a model-agnostic harness—one built on design principles that work across Claude, GPT, Gemini, and Grok without model-specific tuning—can match or exceed those scores.

“Agent performance is really not locked into any single model provider,” Deoras said. “That opens up the opportunity to build a variety of applications without being constrained to a particular model.”

To back the claim, AWS is open-sourcing its framework, called Simple Strands Agent, which the researchers say outperformed popular open-source alternatives across three major industry benchmarks.

The deeper argument underlying all of it is one the industry has been slow to absorb. Most AI performance gains to date, the research argues, are brittle: optimizations that overfit to the quirks of a specific model version, then evaporate when the model improves.

“As models improve, these behaviors change, making such gains brittle and noncompounding,” according to the research.

What’s needed instead are invariant principles—design choices that survive model upgrades because they’re engineered into the harness, not the model. Deoras said the discovery of those invariants was the finding that surprised him most.

“Despite all the differences in modeling philosophy, there is a common invariant property that connects all these models together,” he said. “I didn’t expect that, but this data just naturally emerged from our observability traces.”

The practical implication is pointed for any organization building on AI. The team responsible for re-architecting a harness every time a new model drops—and that is currently every organization deploying agents—is spending its time on the wrong problem.

“The team is overwhelmed by model switching and re-architecting anytime there is a model upgrade,” Deoras said.

The vision he describes for where agents are headed is not one of unchecked autonomy, but of something more considered: humans setting direction, agents executing, and sandboxes catching the errors in between.

“You want humans to be in the driver’s seat to direct the work and then take the hands off,” he said. “That’s the future we are marching towards.”

Whether the industry gets there before flying blind catches up with it is, for now, an open question.