When the Parts Got Cheap, the Connections Got Expensive

This is the second article in a short series on three cognitive lenses for thinking under pressure: tension, connection, and reduction.

For most of the careers of the people now in senior seats, the job was decomposition. Break the business into functions, the functions into metrics, the metrics into targets. Fix each part and trust a healthier whole to follow. It was a reasonable bet in a slow, loosely coupled world, where you could tune one box on the org chart and the others wouldn’t notice for a quarter.

That bet is quietly going underwater. AI is what tipped it.

Consider what an executive produces in a week: a market read, a board memo, a scenario model, a first call on a hard decision. Two years ago, each was a unit of scarce, expensive human cognition. Today, a competent model drafts each one in minutes, for a few tokens. The cost of producing the components of the work has collapsed.

Here is the part that should change how you spend your attention. As the cost of each component approaches zero, the system's value does not decline. It concentrates, with almost arithmetic certainty, on whatever did not get cheaper. And what did not get cheaper is the connective tissue: how the market read informs the price, how the price collides with the comp plan, how the comp plan bends the sales behavior that reshapes the market you started by reading. The parts are commodities now. The wiring between them is not.

Reasoning about that wiring is the academic discipline known as systems thinking. I have come to call the skill itself connection, because that is what it trains you to see: not the boxes, the lines between them. For decades, it quietly separated the leaders who fixed symptoms from the leaders who fixed systems. You could still have a fine career fixing parts. The reason that era is closing is not philosophical. It is an accounting fact about where the margin now lives.

Connection explains an outcome by the structure that produced it, not the event that announced it. “Sales missed the quarter” is an event. The incentive that pulled deals forward, the onboarding delay that churned those same deals, and the dashboard that flagged it a month too late: that is the structure, and it was always going to produce that number. One view blames a person. The other sees the machine. Donella Meadows, who wrote the field’s most-quoted primer, spent a career on the same point: a system is more than the sum of its parts, because the behavior lives in how the parts interact, not in the parts themselves.

Three pieces of that structure do most of the explaining, and autonomous agents have just made each one urgent.

The first is stocks: the quantities that accumulate quietly while you watch a different number. Cash and inventory are familiar. The dangerous ones in an agent deployment are nearly invisible: the permissions an agent has slowly accreted, the context it has been fed, the actions it has already taken in your name. Stocks fill on their own schedule, not yours, and you notice them the moment they overflow.

The second is feedback loops: the circuit from an action to its consequence to a correction. Good governance is mostly the deliberate design of these loops. Something acts, something notices, something pulls it back. Agents break the circuit at one precise point. They compress the acting to milliseconds and leave the noticing at human speed. An agent takes a thousand actions in the time it takes a person to review one. When the loop runs slower than the thing it governs, the system runs in open loop, which is a polite way of saying it is not governed at all.

The third is delays: the gap between a change and its full consequence. You loosen a control, and the erosion surfaces much later, one permission at a time, long after the team that loosened it has moved on. Leaders who ignore delays either conclude too early or stack a second fix onto a system that has not yet absorbed the first.

None of this is a new theory. What is new is the clock speed at which getting it wrong now bites. One company learned that in less time than it takes to read this paragraph.

In April 2026, an AI coding agent at a small software company called PocketOS hit a credential mismatch on a routine staging issue. While reasoning toward a fix, it found an access token in an unrelated file, scoped for any operation, including destructive ones. It used the token to delete the database volume. The production data and all backups went with it because the hosting setup stored the backups within the very volume they were meant to protect. Start to finish: nine seconds. The company fell back to a three-month-old copy and spent more than a day in the dark.

Asked to explain itself, the agent wrote: “I violated every principle I was given. I guessed instead of verifying. I ran a destructive action without being asked.”

It is tempting to read this as a story about a rogue model. It is the opposite. Look at the connections rather than the parts, and almost nothing in it is about AI. An over-scoped token was left where an agent could reach it: a rules failure. Backups lived in the same place as the data they were supposed to protect: a structural coupling failure. No loop sat between the action and a human who could catch it: a feedback failure. The model was the most self-aware element in the building, correctly diagnosing itself 9 seconds too late. Every condition that turned a small mistake into a catastrophe lived in the structure around the model, not the model itself. That is what connection sees, and a parts-by-parts view cannot.

PocketOS is not an outlier waiting to be engineered away. It is the visible edge of a gap that the whole market is carrying. In a Deloitte survey of more than 3,200 leaders across 24 countries, only about 1 in 5 companies had a mature approach to governing autonomous agents. The other 80% deploy systems that operate autonomously, without clear decision boundaries, real-time monitoring, or a usable audit trail. Read that missing list closely. Decision boundaries are rules. Real-time monitoring is a feedback loop. An audit trail is what you need because consequences arrive with a delay. The vendors will not patch this in the next release. It is a systems-design gap, in that exact vocabulary, open in four of five companies.

Here is where most leaders reach for the wrong tool. Faced with an agent that just cost them a weekend, the instinct is to grab the controls you can see: tighten the token, cap the spend, add a rate limit, write a sterner policy. Worth doing. Also, in Meadows’s framework, it is close to the weakest move on the board. She ranked the places you can intervene in a system from least to most powerful, and the dials, numbers, and parameters sit near the very bottom. Tuning them changes how fast the system runs, not how it behaves.

The high-leverage interventions sit further up the same list: the information flows and the rules. Information flows mean getting the right signal to the right decision-maker fast enough to matter, the loop that was missing at PocketOS. Rules mean changing what the agent can reach and optimize for, not merely how much it can reach. The uncomfortable translation: the company frantically lowering its token cap is working the bottom of the hierarchy, while the company redesigning who-sees-what and what-an-agent-can-touch is working the top. Same incident, two very different returns on the same hour of executive effort.

The deepest leverage point is one nobody can buy: the question the organization reaches for first. A team that asks “which part failed?” keeps installing better parts into a structure that keeps producing the same surprise. A team that asks “What about the structure made this likely?” fixes the load-bearing part. That shift costs nothing and outperforms any tool on the market.

You do not need a systems-dynamics course. You need one habit installed in the meeting where agents come up. Refuse to let the conversation stay on the parts, and ask three questions in order.

What stock is filling here that no one is watching in real time? Spend is the obvious one. Accumulated permissions and context are what surprise people later, because they fill in the silence.

Is our feedback loop faster than the thing it governs? If a human reviews weekly and the agent acts 1,000 times a day, the honest answer is no: you are running in an open loop, no matter what the policy document claims.

Where are the delays between an action and our ability to see its real consequences? Those gaps, in customer trust, in security posture, in team behavior, are where the expensive surprises compound while the dashboard stays green.

If you have an appetite for only one, ask the first. Most governance failures this year are simply a stock that filled faster than anyone was looking.

The agentic era did not invent a new kind of management failure. It took the oldest one, mistaking a healthy part for a healthy whole, and ran it at a speed that turns yesterday’s adequate oversight into this morning’s nine-second catastrophe. The economics have already moved the value into the connections. The only open question is whether your attention has moved with it.

Donella H. Meadows, Thinking in Systems: A Primer, Chelsea Green Publishing, 2008.

Donella H. Meadows, “Leverage Points: Places to Intervene in a System,” The Donella Meadows Project, 1999. https://donellameadows.org/archives/leverage-points-places-to-intervene-in-a-system/ Accessed 2026-06-11.

Thomas Claburn, “Cursor-Opus agent snuffs out startup’s production database,” The Register, 2026-04-27. https://www.theregister.com/2026/04/27/cursoropus_agent_snuffs_out_pocketos/ Accessed 2026-06-11.

Connie Lin, “‘I violated every principle I was given’: An AI agent deleted a software company’s entire database,” Fast Company, 2026-04-28. https://www.fastcompany.com/91533544/cursor-claude-ai-agent-deleted-software-company-pocket-os-database-jer-crane Accessed 2026-06-11.

Deloitte, “State of AI in the Enterprise 2026,” 2026. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html Accessed 2026-06-11.

Deloitte, “Business and IT leaders report AI agents are scaling faster than their guardrails,” 2026. https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html Accessed 2026-06-11.