AI Context Engineering: Why More Information Isn't Always Better

The Context Paradox in AI Systems

As AI systems expand their context windows to handle more tokens and larger datasets, a counterintuitive problem is emerging: more context can actually degrade performance. Jake Sortor, Public Sector Strategy Lead at Blitzy, recently challenged the assumption that maximum context equals better AI outcomes, particularly in defense and intelligence applications.

Speaking at TedX MIT Boston, Sortor distinguished between quantitative context—the technical capacity measured in tokens—and qualitative context, which involves how AI systems understand their operating environment. Both dimensions matter, but the industry's focus on expanding context windows may be missing a more fundamental engineering challenge.

Three Critical Failure Modes

Sortor identified three ways that context failures manifest in intelligence systems, all of which apply to AI implementations:

Misclassification occurs when surrounding context makes an anomalous signal appear routine, or vice versa. Assembly failure happens when relevant pieces of information exist but never coalesce into a coherent picture. Update failure means the system continues reasoning from an outdated frame after reality has changed.

These failures characterized historical intelligence breakdowns from Pearl Harbor to the search for WMDs in Iraq. In each case, Sortor noted, information wasn't the problem—the system's ability to properly classify, assemble, and refresh its contextual frame was.

Why Maximum Context Can Hurt Performance

When critical signals get buried in long documents, when dissenting data is smoothed away by summarization, or when systems produce confident answers that hide underlying uncertainty, long context becomes a liability rather than an asset. Sortor warned that relevant information surrounded by distractions can cause performance degradation, even as the industry races to expand context windows.

"The goal is not maximum context," Sortor emphasized. "The goal is optimized context." This requires intentional design choices about what gets retrieved, what gets filtered, what dissent survives, and what uncertainty reaches human operators.

Why it matters

As AI systems move into high-stakes defense and intelligence environments, treating context as an automatic byproduct of larger models risks "executing historical intelligence failures at the speed of compute." The next competitive advantage won't come from feeding AI more information, but from deliberately engineering the context that transforms raw data into correct meaning—a capability Sortor characterized as a national security imperative.

Context as Engineering Discipline

Sortor argued that context engineering must become a deliberate practice, not an afterthought. The fundamental question isn't how much information a system can hold, but whether the architecture provides enough context to make signals meaningful without burying them in noise.

In defense environments defined by fragments, time pressure, and competing interpretations, better data fusion and stronger models matter—but they don't answer the harder engineering question: What should the model see, in what form, at what moment, and for what purpose?

These insights were first reported by John Werner in Forbes, based on Sortor's presentation at TedX MIT Boston.