AI Myopia Gets More Dangerous as Agents Gain Autonomy

In March, I wrote about a class of AI failure I called Context Myopia.

Context Myopia is when an AI system can reason correctly and execute a task well, yet still produces work that is wrong for the company because it was missing information it didn’t know it needed.

At the time, I argued that this failure would become more important as models improved, and that the model was becoming less of the constraint. The systems around it, and the management practices that determine what agents know before they start working, were becoming a larger part of AI quality.

Six months later new research is showing how broadly that shift is taking hold.

Redis surveyed IT and AI infrastructure leaders for its 2026 State of Context Engineering report. A few findings stood out:

• 69% say an agent that cannot navigate relationships across systems creates more operational risk than one that hallucinates.
• 73% say agents fail more often because of broken context than broken models.
• 79% consider context very important to their AI and data goals.
• 81% are still managing it through ad hoc or exploratory approaches.
• 69% of the most mature organizations still feed critical decisions with stale data.

Among the organizations Redis considers most mature in context engineering, 97% rank a failure to navigate relationships across systems a greater risk than hallucination.

That is a meaningful change in how people building these systems think about AI reliability.

Better AI makes myopia harder to see

Hallucination became the defining AI quality problem because the failure was often visible in the output. Fabricated facts, sources, and citations triggered a certain “smell” once users of AI tools became familiar with the pattern. Furthermore, facts and sources could be directly checked by an adversarial reviewer in a somewhat reliable manner.

Myopia behaves differently because the output can be accurate given the input to the isolated task. The work appears technically strong, and surface-level checks validate quality and attempt to detect signs of error. But with myopia the source of the error is deeper and harder to detect. Instead of a miscalculation, it’s a decision, dependency, customer commitment, or change that altered what the company actually needed.

That was the problem in the example I used in my March blog post. An agent was asked to add access control to a product, and it implemented OAuth correctly and soundly. But unbeknownst to it, the team had already decided to use an alternate security technique, SAML. The implementing agent had no access to that decision so it made a reasonable assumption. And reviewing the code itself would not reveal the problem either because the code was fine. The agent was correctly solving the wrong version of the task.

Myopia and related errors become more consequential as AI improves, because better execution creates stronger evidence that the work is trustworthy. When an agent produces clean code, coherent analysis, or a polished recommendation, the quality of the output gives the reviewer fewer reasons to question the premise underneath it.

A more capable model can therefore make myopia less obvious. We even see it in the human world.

Think of a good doctor seeing a patient with a dry cough that has lasted for a few months. No fever, lungs clear, nothing on the chest X-ray. The doctor works through the usual causes, concludes asthma is likely, and explains to the patient the prognosis and prescribes an inhaler. It’s a textbook diagnosis given the data. But what the doctor didn’t see was that another physician started the patient on a certain type of blood pressure medication a few months earlier, and persistent dry cough is a common side effect of that particular drug.

In this example, the doctor’s reasoning was right. The explanation was convincing and correct. But the input (the chart) was incomplete, and so was the diagnosis. And to make it worse, if the doctor now writes “asthma” into the chart, future physicians will be biased towards that diagnosis as well.

Autonomy changes the cost of myopia

The other change since March is how quickly the market is moving from AI that produces an answer toward AI that acts.

Anthropic describes the current shift in practical terms. Agents can write and execute code, manage files, and carry out tasks across multiple applications. That autonomy brings increased productivity and speed. But as agents operate with less human oversight, there is more opportunity for them to misread intent and take unintended actions.

Despite the shortcomings, companies expect that autonomy will continue to increase quickly. Deloitte surveyed 501 sr. managers and C-suite executives and found that 61% expect most of the AI agents they use to become generally autonomous within four years, with humans primarily providing oversight. Yet only 5% say their business processes are highly prepared for agents, and just 15% have scaled orchestrated, cross-functional multi-agent adoption.

In the past and present, with AI acting as a copilot instead of fully autonomously, a context failure often ends with a human correcting a draft, providing additional steering, or even completely redoing a task. There is a cost, both in time and tokens, but the risk is mitigated because of human review and interception.

Autonomous agents operating on incorrect information not only impact the initial task, but downstream tasks and peer agents as well. The more capable the agent becomes at moving work forward, the more important it becomes that the agent knows whether the work should be moving in that direction at all.

How do we evaluate AI output quality in a world of myopia?

For the last several years, much of AI quality work has focused on what comes out of the model given input. Hence the rise of “evals”. Teams evaluate accuracy, test generated code, add guardrails, monitor behavior, and put humans in review loops. Those controls still matter, but they are also weakest against failure where the missing information and malformed task definition cause errors at a more systemic level.

It is no longer sufficient to ask whether an agent executed a task correctly given its input. The system also needs to establish whether the agent had enough knowledge of the company to understand the task correctly in the first place. That includes whether its information is current, whether a relevant decision has changed, and whether something elsewhere in the company changes what the work means or intends.

This is why context becomes more than a retrieval problem as agents become coworkers. A search system can surface relevant information, but it cannot guarantee the agent knows what it needs to look for. Myopia often starts with an unknown unknown, where the agent has no reason to search because nothing in its field of view tells it that anything is missing.

A seat at the table matters more when the AI can act

In March, I argued that AI coworkers need the same access to decisions and background that helps human employees do good work. I believe this more than ever.

We are starting to see a rise in more “embodied”, persistent AI entities that have more autonomy and more capacity to operate on their own. In some cases, like Kinelo’s AI Coworkers, they even go as far as feeling like remote coworkers with job descriptions, roles, and the ability to function effectively within a human team.

With this rise in agency of AI systems, the cost of Context Myopia and related errors is increasing quickly. When AI mainly produced drafts, Myopia created rework. As AI increasingly acts on its own output, failures can spread through the team and its work before anyone realizes the issue.

A seat at the table was an AI quality issue six months ago. As agents gain authority to act, it is becoming an operating requirement.