Why the harness matters far more than the model for super intelligence success

Federal defense and intelligence teams are under tremendous pressure right now from the top down to demonstrate that super intelligence (SI) can and will make a difference in their operations.

Part of that pressure comes from not having a clear directive much beyond “figure out how to do SI.”  Even where there is a vision for using SI, many agency leaders are still grappling with how best to deploy SI across existing systems, domains and workflows.

I regularly encounter scenarios in my work with agencies that go something like this: A deployment team sifts through a barrage of vendor claims that their AI-capable tools will solve a particular problem. They settle on an approach and put it to the test. On Monday, the AI tool feels almost miraculous. It analyzes complex threat streams, drafts technical code, and recalls vast historical data. Yet on Friday, that same tool fails to retain context across three basic steps, hallucinates facts, and generates nonsensical answers. Overnight, enthusiasm turns to deep skepticism. 

What many agencies are still learning is that the same AI model, with the same built-in algorithms, can yield significantly different operational experiences. This poses special challenges for defense and intelligence agencies, where trusted answers are critical.  

To better understand what’s going on under the hood, it’s important to remember that at their core, large language models are probabilistic engines. Every output is the result of rolling the equivalent of mathematical dice.

What is often misunderstood is that discrepancies can reflect more than the model itself. SI engines depend on an underlying harness — the software infrastructure and scaffolding that includes terminal commands, code interpreters, context and memory architecture, execution instructions, governance guardrails, and integrated tools that surround the engine. Harnesses orchestrate how data, tools and controls work together across mission systems and operating environments, helping preserve context, enforce access controls and maintain human oversight as work moves between tools.

Most importantly, harnesses rely on machine-readable datasets used throughout your ecosystem that models need to understand to complete their requested tasks.

One factor often overlooked by defense and intelligence agencies is that language models do not intrinsically understand physics-based sensor data used across defense and intelligence streams. Nor can standard retrieval systems discern why 20-minute-old tactical intelligence carries vastly different operational weight than 20-year-old static records.

That’s why the data AI models used by defense and intelligence agencies must be properly formatted, so that machines understand what the data (or object) represents, how they know and can verify that, and why it matters.

If I have data indicating there’s a threatening actor on a ship in the Pacific Ocean, but I don’t know where that information came from, how old it is or its value, I don’t have enough intelligence to make a useful decision.  

For defense and intelligence agencies, unreliable outputs can have serious operational consequences. Leaders need to understand the evidence behind an output, its limitations and when human review is required before action.

At the same time, the reality is:

· Not every algorithm is built for battle.

· Not every dataset is machine-ready.

· Not every intelligence output warrants automated action.

Once leadership recognizes this reality, the foundational requirement becomes clearer: super intelligence is entirely dependent on data structure. Throwing billions of dollars and massive GPU clusters at frontier models yields diminishing returns if the underlying data substrate is broken.

Beyond the data regimen, agencies must take responsibility for building their own orchestration scaffolding before selecting commercial models. Rather than locking into a single vendor’s proprietary software stack, leaders need to establish a model-of-models architecture managed by a dedicated integrator that coordinates specialized tools, memory, context protocols, and governance guardrails.

Finally, while frontier models offer unprecedented speed and scale, traditional expert systems remain critical for delivering strict logic, auditability, and deterministic rules. Fusing these approaches in a neuro-symbolic AI layer — that is, integrating the probabilistic pattern recognition of deep learning neural networks with rule-based logical reasoning — is key to providing the exact traceability agencies require for high-consequence decision-making.

By applying attribute-based access control (ABAC) down to the cellular level, the system automatically resolves target representations across security domains. A warfighter operating on a Secret network receives precise, actionable location data on an adversary vessel without exposing the classified sources and methods that are visible in Top Secret enclaves. That is the power of a proper data harness: raw intelligence transformed into traceable, defensible operational decisions.

Eighteen years in defense and intelligence have reinforced a practical lesson: SI investments must be evaluated in the context of the mission systems, data and workflows they need to support.

For leaders, the next step is to examine an existing SI effort: Where does it lose context? Can users trace its outputs to supporting evidence? Are access controls and human oversight applied consistently? Can it work across the environments the mission requires?

Addressing those gaps can help agencies get more value from existing investments and move from isolated demonstrations toward governed, reliable SI at operational scale.

In fact, I recently made a wager with an enterprise customer. I bet him that if he stopped funding yet another commercial software license and redirected those dollars straight into fixing his core data structure, his SI performance wouldn’t just improve—it would soar by orders of magnitude by the end of the year.

For defense and intelligence leaders intent on truly leveraging SI, look beyond the models. Understand what actually happens at the orchestration level. But most of all, fix the data foundation, and the performance will follow.

Learn more about how SAIC is helping civilians, defense and intelligence leverage Super Intelligence.