Divergence in the Human Loop: Part 2

Restructure for AI

In Part 1 of this series, Human in the Loop, we shared our view that the deployment of hundreds, even thousands of agents on systems has fractured the traditional control structure of companies. Humans are unable to keep pace with the requirements of a well-managed, and closely-supervised domain. We presented a summary of three remedies to ensure humans are not only "in the loop", but effective for their assigned role. In Part 2, we go a little deeper in describing our experience with implementing the 'decision design remedy' as described by Margaret Mitchell in her position paper. Of the three, decision design is where the enterprise itself has to change.

Every Decision Point, Revisited

Decision design treats expert attention as the scarcest resource a company has, and spends it only where the stakes justify it.¹ Applied seriously, that reaches far beyond software. Consider purchasing. OpenAI and Stripe's Agentic Commerce Protocol lets an agent complete a purchase with a payment token scoped to one merchant and one cart total, without ever seeing the card.² That solves the payment rail. It does not answer the governance question: when an agent holds the company card, who decides what? The answer is not more approvals. A routine reorder from an approved vendor, inside a budget, clears on its own and lands in the record. A new vendor or an amount above threshold needs a one-line justification. A contract, a capital item or a recurring commitment stops for an accountable owner. The control is the rating rule, the mandate (scope, budget, expiry) and the record — not the number of clicks. Every approval chain in the company deserves the same re-examination.

Complexity Is the Quiet Risk

Control is half the story. The other half is that systems left to change drift toward complexity. Dr. Meir Lehman observed in his 1980 landmark treatise on the laws of software evolution: a system's complexity increases unless work is done to reduce it.³ More than that, he authored the maxim known as Lehman's Uncertainty principle, a collection of laws which concludes that "no system can ever be relied upon to be correct". In other words, complexity is rarely correct. Think of this way: AI Agents are prone to accelerate the drift in systems, because adding is easier than reusing. GitClear's analysis of 211 million changed lines found that refactoring fell from 25 percent of changes in 2021 to under 10 percent in 2024, while copy-pasted code rose.⁴ Fred Brooks separated the essential complexity of a problem from the accidental complexity we pile on top.⁵ Agents can generate the accidental kind at machine speed, and every unnecessary layer is operating cost for years. A control system that only asks "who approved this?" misses the better question: "what did this make harder?" Or the question we like to ask: "How do we simplify this?".

Oversight at the Edges of Change

The Accelerator, our platform for crews of agents that build and maintain web applications, places human review at the edges of change, where complexity invades most forcefully.

A complexity rating on every work order. Before work starts, the team lead (an AI agent designed for that role) rates each request (from agents and humans) by what is at stake and how much expert effort it needs. Payments, personal data, sign-in and security are always rated high, and if a build reaches a protected area the rating can only rise. The rating sets the checkpoints: none for a low change, one brief for a medium one, and design, plan, preview and release for a high one.

A test bundle before any code moves. Every step adds its own unit and integration tests. Nothing advances until the type check, the full suite and a production build pass on the exact candidate, and the merged result is checked again before release. Each result is bound to the commit it judged, and the record is written into the owner's own repository. When we added Stripe checkout to a test store this week, the owner made five decisions while the suite grew from 243 to 276 tests.

Simplicity as a standing rule. The crew makes the smallest change that works, reuses existing code before adding layers, and never refactors outside the agreed scope. A code census at every proof counts functions, components, routes and the most branch-heavy files, so drift shows up as a number before it shows up as cost.

Managing Complexity, Not the Code

Before agents, oversight meant a person reviewing every line. That model cannot survive machine speed, and it was never where the value was. The human role moves to the edges: deciding what is worth doing, judging the evidence, and refusing complexity the business does not need. The organizations that win with agents will not review more code. They will govern decisions, and complexity, by design.

So rather than divergance in our operating model, we have both symmetry and convergence on a model with humans in the loop, and agents in the mix. Give us a call, and lets discuss how our accelerator platform can help your company deliver with AI.

At Strategic Machines, we deliver the Accelerator: an expert crew of AI agents, trained for your industry, that improves your digital products under your direction. Every change arrives as a defined work package — the request, the design, the plan, the tests, your approvals and the release — kept together as a lasting record you own.

See it in practice: a shop owner adds Stripe checkout to her store, from request to release, with every decision on the record.

Let's talk.

SOURCES AND REFERENCES

¹ Margaret Mitchell, Avijit Ghosh and Samir Passi, "AI Agents Push Humans Out of the Loop," arXiv:2608.23642 (2026). Decision design is one of the paper's three remedies, with strategic friction and behavioral monitoring.

² Stripe, "Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol co-developed with OpenAI" (2025).

³ Meir M. Lehman, "Programs, Life Cycles, and Laws of Software Evolution," Proceedings of the IEEE 68(9), 1980.

⁴ GitClear, "AI Copilot Code Quality: 2025 Look Back at 12 Months of Data" (2025).

⁵ Frederick P. Brooks Jr., "No Silver Bullet: Essence and Accident in Software Engineering," Computer 20(4), 1987.