AI Agents Don't Inherit Tech Debt—They Weaponize It

I spend half my week talking to enterprise teams who want to deploy autonomous AI. They watch a slick demo, see a cursor move by itself on a screen, and immediately want it integrated into their production environment.

Then reality hits.

Look at what is actually happening in the market today. Microsoft is actively testing autonomous AI assistants for the daily workday. At the exact same time, we have competing philosophies on how to keep these things from burning down the house. Cisco is pushing a narrative with their Astrix deal that AI agents don’t actually need extra security layers. Their argument? Standard enterprise security protocols should suffice.

I call bullshit.

If your underlying Identity and Access Management (IAM) is a mess — and let’s be honest, whose isn’t? — an autonomous agent doesn't just inherit your technical debt. It weaponizes it. There is a reason Apple is intentionally building AI agents with hard limits. You do not scale autonomy without first scaling constraints. A thousand autonomous actions a minute compounding on broken permissions isn't innovation. It’s an order of magnitude increase in your blast radius.

The Geopolitical Cap Table

Building from Singapore, you learn early that software might be eating the world, but geography still dictates the menu.

We like to pretend open-source code and global APIs make borders irrelevant. They don't. Capital is the ultimate border. We just saw this play out with the "Manus Warning" — Beijing's $2 billion blockade on Western investors in Chinese AI.

This isn't just a headline for macroeconomists. It is a direct operational constraint for anyone building in APAC. If you are relying on cross-border capital flows or shared foundational models between the East and West, your risk profile just changed overnight. You have exactly three domains that demand your focus in life: your career, your family, and your finances. Spending your career building on a geopolitical fault line is a massive misallocation of time. The infrastructure layer is splitting. You have to pick a side and build for that reality, or you will get crushed in the middle.

The Feature Trap and Developer Atrophy

Every vendor is racing to add sensory inputs to their agents. Runway just shipped API capabilities that give their characters "eyes" via camera and screen sharing.

It makes for a phenomenal Twitter demo. But how does that actually deploy in a secure enterprise environment? It doesn't — not without months of painful compliance reviews.

We are confusing capabilities with productivity. We see this acutely in our own engineering teams. There is an ongoing AI agent arms race in IDEs, and it is actively sabotaging team creativity. When your editor writes the boilerplate, suggests the logic, and auto-completes the architecture, your developers stop thinking in systems. They become passive reviewers of mediocre machine output.

The compounding effect of developers outsourcing their problem-solving muscle is a net-negative for your engineering culture. Are we building tools that make our engineers faster, or tools that make them lazy? Orders of magnitude improvements in output don't come from auto-completing bad code faster. They come from deep, uninterrupted technical thought.

The Infrastructure Reality

Compute is the ultimate currency. Demos ignore latency; production environments live and die by it.

Look at the hardware moves happening beneath the software noise. Intel and Google are deepening their AI infrastructure alliance around Xeon and custom IPUs. Why? Because running bloated models at scale is destroying enterprise margins.

Even Microsoft is feeling the pinch, quietly launching MAI-Picture-2-Environment friendly, explicitly marketed as a cheaper, faster image model. When the hyperscalers start pivoting to efficiency over raw parameter count, you should pay attention. You cannot build a sustainable business if your unit economics are tied to a model that costs a dollar per inference. Optimization isn't a nice-to-have — it is the only way your startup survives the next 18 months.

The Bottom Line

We are entering the deployment phase of AI. The hype cycle is dead, and the tourists are leaving. What remains is the hard, unglamorous work of integration. If your AI strategy relies on unconstrained agents, geopolitical stability, or the assumption that more "smart" features automatically equal better output, you are going to lose. Constrain your agents. Audit your dependencies. Protect your team's ability to actually think.