First-principles systems thinking
Data was free to copy. Turning it into intelligence never was.
Information costs energy to make useful. Copying data is near-zero cost. Every step that transforms raw data into context, structure, and actionable intelligence requires work. In 2015, I wrote that transformation was the hidden cost behind every information product. The cheaper information became to copy, the more the cost shifted to what you did with it.
Agents have confirmed that thesis at scale. Every inference call has a price. The model assumes nothing, ingests context, and probabilistically matches inputs to outputs, more expensive than a deterministic rule that already knows the answer. What’s also become clear: piecemeal context (loading only what each agent needs for each task) creates a ceiling on what agents can accomplish together. An agent synthesizes only what sits inside its context window. Specialists multiply. Synthesis across them still requires humans.
The minimum energy state for any information system is a priori knowledge: everything pre-classified, rules predetermined, routing deterministic. That’s the floor. Below it, deterministic systems are cheaper, faster, and more reliable than agents reasoning from scratch. Above it, agents earn their cost, handling cases the a priori rules don’t cover.
The floor is defined by context. What we can predetermine is bounded by what we know in advance. What looks like an exception today may be common enough to classify tomorrow. What’s deterministic in one organization may require agent reasoning in another. The practical answer: guess at the line, build systems that let you observe where it actually falls, and correct over time.
At AWS Health, we codified this into a maturity framework: here are the decisions about service health that can be predetermined at each maturity stage. Define those first. Build deterministic routing for them. Then bring agents in for the reasoning your infrastructure can’t yet handle.
Going straight to agents without that foundation is like hiring a PhD graduate to hand-deliver water to every person in a building instead of installing pipes. The PhD is capable of far more. Pipes are unglamorous. But without the pipes, you’re paying for judgment on a problem that infrastructure solved a century ago. Building the pipes is a discipline in itself, and the subject of the next article.
The PM question is: what do we know well enough to predetermine, and what still requires judgment? Agents made the cost of getting that answer wrong concrete and measurable.
Human-in-the-loop processes have always existed. Approval chains were the original mechanism for catching what the deterministic path misses. As agents accelerate the decision cycle, the guardrail design accelerates with it. Reviewing for second-order effects and unintended consequences requires agents with progressively broader context at each review layer. Designing that adversarial review layer is the frontier after that.