The Real Bottleneck Is Not Money, It Is Remembering What You Are Trying to Do

Hatched by Noah

Jul 29, 2026

10 min read

1 views

91%

The Strange Common Problem Behind Broken States and Smart Machines

What do California’s housing crisis, its pension mess, its homeless encampments, and a coding agent’s need for a Todo list have in common?

At first glance, almost nothing. One is public policy, the other is software engineering. Yet both revolve around the same failure mode: a system that keeps adding more activity while losing sight of the outcome it was supposed to produce.

That is the hidden connection. California does not merely spend too much. Claude does not merely need more tools. In both cases, the deeper problem is action space without accountability. When a system has many possible moves but no disciplined way to measure whether those moves are working, it drifts toward complexity, process, and self-justifying motion. It becomes very good at appearing busy and very bad at finishing the job.

That is why these two worlds rhyme so eerily. Sacramento adds bills, fees, consultants, and layers of review the way an agent adds tools, reminders, subagents, and search paths. Both are trying to solve a legitimate problem. Both often make the same mistake: confusing more structure with better judgment.

When Institutions Forget Their Output, They Worship Their Process

California’s story is not best understood as a shortage story. It is a story of process inflation.

Spending rises 75 percent. Outcomes stay flat or get worse. A high-speed rail project absorbs billions, years of litigation, armies of consultants, and still struggles to resemble a finished product. Housing becomes so burdened by fees, codes, and lawsuits that even starter condos turn into fantasy objects. Home insurance stops functioning because risk is suppressed rather than priced. Homelessness persists not just because poverty exists, but because the system can neither prevent vulnerability nor move people rapidly into shelter and treatment.

The pattern is not hard to see once you name it: every failure generates a new rule, every rule generates a new layer, and every layer creates new stakeholders who now need the layer to remain in place. The system grows not because it works, but because it can justify its own continuation.

This is where the analogy to agent design becomes useful. A coding agent with only a vague objective, such as “fix the app,” will often wander. So it gets a Todo list. Then it forgets the Todo list, so it gets reminders. Then subagents are introduced, so coordination becomes a problem, so Tasks replace Todos. Then the model cannot remember how to use the product it is operating, so progressive disclosure and a guide subagent are added.

At first, all of this seems like progress. And it is, up to a point. But every new tool also expands the agent’s action space. That means more freedom, yes, but also more room for confusion. Without careful boundaries, the agent spends more time deciding how to decide than deciding.

That is exactly what bureaucracies do.

A system does not become effective by multiplying its options. It becomes effective when every option is tied to a visible outcome.

California’s problem, then, is not simply that it has bad intentions, or too little revenue, or too many enemies. It is that the state has built an elaborate machine for producing motion without a reliable mechanism for asking a blunt question: Did this actually help?

Safetyism, Like Over-Tooling, Creates the Illusion of Seriousness

One of the most revealing threads in the California picture is the logic of safetyism. After every bad headline, the instinct is to add a rule. Something happened, therefore another safeguard is needed. Another check. Another balance. Another fee. Another compliance step.

That sounds responsible. It often feels moral. But it is frequently a way of avoiding the harder task, which is to make tradeoffs and own results.

The same thing happens in software systems. When an agent fails, the instinct is not always to simplify the environment. More often, engineers add a new tool, a new reminder, a new retrieval layer, or a new subagent. Each addition seems rational in isolation. But when the model keeps missing the point, the answer is usually not another instrument. The answer is a better way to preserve focus.

California’s governance has increasingly rewarded the performance of care rather than the delivery of care. That distinction matters. If officials can signal concern by layering on process, they can avoid the embarrassment of measurable responsibility. A bill can be passed. A committee can be formed. A report can be commissioned. A framework can be announced. None of these things requires the system to prove that the lives of ordinary people improved.

That is why so much public policy becomes a theater of moral seriousness. Everyone can claim they are working on housing, schools, homelessness, climate, affordability, and equity all at once. But if the metrics do not move, the moral language becomes camouflage for institutional inertia.

The lesson from agent design is harsh but clarifying: more capability increases the need for more disciplined prioritization. The smarter Claude becomes, the more important it is that it can build context selectively, not indiscriminately. Likewise, the more powerful a government becomes, the more dangerous it is when it loses the ability to choose among priorities.

This is why the mantra “we need more revenue” is so often incomplete. Revenue is a fuel. It is not a steering wheel. If the destination is unclear, more fuel simply gets you lost faster.

The Best Systems Do Not Add Power First. They Add Feedback First.

The most interesting counterexample in the California story is not a theory. It is a method.

Public dashboards. Baselines. Targets. Audits. Outcome tracking. Zero-based budgeting. Publicly visible metrics. These are not glamorous reforms, but they are the hidden infrastructure of competent systems.

A useful mental model here is to think in terms of closed loops versus open loops. In an open loop, a system acts and hopes. In a closed loop, it acts, measures, adjusts, and repeats. The difference is everything.

Claude Code evolved in exactly this direction. A simple note to “remember your task” was not enough. So the system got a Todo list. Then reminders. Then Task-based coordination. Then progressive disclosure for context discovery. Then a guide subagent for the narrow case of asking about itself. Each step was not just more complexity. It was an attempt to close a loop that had been open.

California needs the same transition. Right now, too many government processes are open loops. Funds are allocated, programs are launched, and the feedback is weak or absent. Audit recommendations sit unread. Agencies are judged by activity, not outcomes. Legislators are rewarded for bill volume, not civic improvement.

The consequence is predictable. If you do not measure the outcome, you eventually measure the ritual.

This is why a focus on basics matters so much. Lower the cost of housing. Lower the cost of energy. Improve schools. Make neighborhoods safer. Reduce unsheltered homelessness. Maintain roads. Those are not separate hobbies. They are the core operating system of a livable state. When they fail, everything else gets more expensive, more fragile, and more politically toxic.

The strong version of this thesis is uncomfortable but important: many policy debates are not really about ideology. They are about whether a system is willing to let evidence overrule identity.

Good intentions are not enough when the feedback loop is broken.

Why Good Intentions Keep Failing in Both Politics and AI

There is a reason both systems, government and agents, can become overcomplicated while still feeling underperforming. The people inside them are usually not stupid, malicious, or indifferent. They are often trying to be helpful.

That is what makes the failure so persistent.

In California, the road to dysfunction is paved with sincere motives. Environmental protection becomes endless litigation. Compassion becomes permissiveness when addiction and psychosis are treated as though they should never be forcibly interrupted. Safety becomes process for process’s sake. Equity becomes a slogan that can survive even when the working class is getting crushed by housing and energy costs.

In a coding agent, the parallel failure is similar. More tools can feel like more intelligence. More search can feel like more rigor. More reminders can feel like more reliability. But unless the system is built to preserve focus, every new affordance becomes another opportunity to wander.

This is the central insight: competence is not the same as breadth.

A competent governor does not necessarily try to solve every problem directly. A competent agent does not necessarily use every tool available. Both need to know what not to do. Both need to know what outcome matters most. Both need to accept that the fastest route to failure is often attempting to be useful in too many ways at once.

That is why the best institutional reforms are often subtractive. Remove a fee. Cut a layer. Shorten a timeline. Narrow the scope. Price risk honestly. Shut down low-value programs. Defer less, decide more. In software, the equivalent might be removing a tool, replacing an all-purpose mechanism with a more selective one, or limiting context instead of flooding it.

A system that cannot say no is usually not compassionate. It is ungovernable.

What California Would Look Like If It Learned to Think Like an Agent

Imagine if a state were designed the way the better agent architectures are designed.

First, it would have a small number of explicit goals. Not 100 priorities, not 12 overlapping visions, but a handful of outcomes that matter most. Housing supply. Cost of living. School performance. Safety. Homelessness reduction. Infrastructure reliability.

Second, every goal would have a visible metric. Not symbolic metrics, but numbers the public can actually inspect. How many homes got built? What did they cost per square foot? How long did permits take? How many people moved from street to shelter to stable housing? How many audit recommendations got implemented?

Third, the system would make it easy to search for root causes before adding new interventions. In agent terms, that is the difference between blindly guessing and using context well. In government terms, it means asking why a project fails before launching a fresh program on top of the failure.

Fourth, the system would treat coordination as a first-class problem. California often behaves as though there is one main villain, but the more accurate diagnosis is a swarm of incentives. Contractors, lawyers, unions, agencies, advocates, elected officials, and donor networks all push in rational ways. The problem is not that one actor steals the whole pie. It is that too many actors are paid from the friction generated by the pie being hard to serve.

That is why reform is so difficult and so necessary. You are not just fixing a policy. You are changing who benefits from dysfunction.

This is exactly why the most valuable political language in the California conversation is not left versus right. It is process versus outcomes.

Key Takeaways

Measure outcomes, not motion. If a system is busier but not better, its process has become the product.
Subtract before you add. When performance slips, ask what can be removed, simplified, or narrowed before introducing new rules or tools.
Treat feedback as infrastructure. Dashboards, audits, and visible baselines are not decorative. They are what make accountability possible.
Do not confuse compassion with permissiveness. In both policy and AI, refusing to intervene can look humane while producing worse outcomes.
Beware of action space without prioritization. More tools, more bills, and more programs all create more room for drift unless there is a strong mechanism for focus.

The Deepest Lesson: Scale Without Focus Is Just Confusion at Higher Cost

The most surprising thing about these two seemingly unrelated stories is that they converge on the same principle: the size of a system matters less than whether it remembers its purpose.

California has become so elaborate that it sometimes cannot answer the simplest question, which is whether the average resident is getting a better life for the money they pay. Claude Code, meanwhile, can become more capable by expanding its tools, but only if those tools help it stay oriented toward the task rather than get lost inside it.

That is the real intellectual crossover here. Institutions, like intelligent systems, do not fail only because they lack power. They fail when power outruns clarity.

The future belongs to systems that can do something harder than expanding. They can focus.

And perhaps that is the most valuable reframe of all: the goal is not to build the biggest apparatus, public or digital. The goal is to build a system that still knows, after all the complexity is added, what it was supposed to accomplish in the first place.