**Cato:** A reservation was not a service. Welcome to the Cato Biannual Digest for April–September twenty twenty six. Over this half-year, AI made it cheaper to produce code, route tasks, and book future computing capacity. But three questions kept coming back. What counts as finished work? Who profits from governing the work an agent does? And when does booked capacity become a service someone can use? **Layla:** Those sound like business questions until you meet the people at the other end. A maintainer gets the patch. A lawyer gets the filing. A household gets the grocery delivery. Each has to decide whether the result is worth keeping—and what happens if it fails. **Cato:** My read is that completed, accepted work—not generated output—is the useful productivity unit. In a test of whole-repository migrations, only five point four percent of cases preserved observable behavior. An agent could produce a patch, and that patch could still fail the test that matters to the receiving team. **Layla:** What does “behavior preserved” mean when an agent changes a whole repository? **Cato:** It means the changed software still does what could be observed before the migration. Not merely that the new code looks plausible or passes one build step. A repository has callers, edge cases, and assumptions tucked into places the person requesting the migration may not know to name. **Layla:** Even if it runs? **Cato:** Even if it runs. A checkout page that loads but calculates shipping incorrectly has not survived the move. The maintainer has to find that difference before customers do. Here’s the anomaly: more drafts need not mean more finished work; the receiving team’s test—not the producing tool’s output count—determines what the institution can use. **Layla:** Cato, I agree that maintainers own what ships. But a whole-repository migration is an unusually hard coding task. I would not carry that five point four percent into every bug fix or small feature. And in law, the review is not evidence that the tool failed. Cooley kept lawyers responsible for AI-drafted S-one materials because a filing needs an accountable professional. **Cato:** Fair. The three cases are not interchangeable. But the buyer still needs a measure downstream of draft production. Uber saw no clear productivity lift from its internal coding tools. That does not mean nobody wrote faster. It means the organization could not yet point to a clear gain after the work passed through its normal process. **Layla:** Then which task categories show review cost falling per accepted result? A lawyer might use a first draft to spend less time assembling standard language and more time checking a risky claim. A maintainer might accept a narrow patch in minutes. Those are gains, even though someone still signs off. **Cato:** That is the strongest counter to my position. Better evaluation and narrower tasks could make review cheap. I had been treating acceptance as the likely enduring scarce input. I need to narrow that: acceptance is necessary; it does not have to stay expensive. We should compare review hours and reopened failures per accepted task, while also counting how many tasks teams take on. **Layla:** Because a team may use the time it saves to attempt work it used to leave undone. If review hours rise in total while accepted work rises faster, the tool may still be valuable. And if defects reopen months later, a fast sign-off was not the whole cost. **Cato:** Right. The falsifier for my stronger claim is concrete: across comparable deployments, review time and reopened failures fall faster than task volume grows. Then this looks like a temporary cost of learning how to use the tools, not a permanent choke point. For now, a vendor reporting suggestions generated is answering a different question from the one the buyer asked. **Layla:** And the person doing the checking needs a real right to reject the result. Otherwise “human in the loop” can mean a worker inherits a stack of machine drafts and the blame for anything missed. That is not a productivity measure. It is a transfer of responsibility. **Cato:** If producing the work is getting cheaper, who decides which result is safe to keep—and which model should do the next task? **Cato:** My read is that task-level allocation is becoming a real product. AT&T reported a fifty six percent reduction in coding cost with a two percent performance decline. That is not a reason to use the cheaper route everywhere. It is a reason to build a workflow that can choose where that trade is acceptable. **Layla:** At what cost? **Cato:** At the cost of measuring the task, setting a quality floor, and handling exceptions. Microsoft reported a higher CyberGym score at nearly half the cost through routing. Those results suggest there is value in choosing a model for the job instead of making one model the answer to every job. **Layla:** There is another split here. Developers could keep Claude Code’s working interface while changing the model underneath it. So model choice and ownership of the workflow are not the same thing. A team may love the place where it sees files, approvals, and past decisions, then swap the engine when price or quality changes. **Cato:** Which is why the control layer has a commercial case. It can set permissions, select models, record what happened, and stop an action that crosses a boundary. OpenRouter’s reported routing scale shows demand for a neutral place to make at least some of those choices. **Layla:** Demand for routing, yes. Not proof of margin. Atlassian already holds a work record and is opening it to agents. It starts with the project history, administrators, and the people who will receive the result. Why pay a separate gateway to recreate that context? And we have not seen gateway loss rates or the contract terms that decide who absorbs a failed action. **Cato:** That is the best counter: an incumbent can bundle controls with the place work already lives. A neutral operator has to offer something the incumbent cannot, or cannot offer credibly across competing systems. **Layla:** If a router chooses a cheaper model, what remains the buyer’s responsibility? **Cato:** The buyer still defines acceptable work and the authority it grants. It cannot outsource every judgment to a routing rule. But the operator choosing the model may be asked to preserve evidence of that choice, enforce the rule the buyer set, and help recover when a result goes wrong. The unusual part isn’t that models can be switched; it’s that the party choosing them may also have to preserve the evidence and handle the exception when that choice fails. **Layla:** Take a purchase order. The agent had permission to draft it, the router logged the model choice, and the amount matched a form. Then the vendor ships to an old address. An audit trail tells us what happened. It does not settle who pays for the shipment. If the gateway sold “governed action,” the buyer will ask what that promise covered. **Cato:** And if the gateway only sold routing, it may say the buyer owned the address book and approval policy. That boundary can be written into a contract. But support, disputes, fraud, and recovery still have costs. I withdraw any inference from OpenRouter’s traffic scale to durable margin. Scale establishes use, not loss-adjusted profit. **Layla:** What can the buyer take when it leaves? Context, policies, evaluation history, logs, and the record of past failures should all be on the list. If those move cleanly and the incumbent offers equivalent controls at lower loss-adjusted cost, the standalone moat is weak. **Cato:** My case for neutral routing survives in narrower form. A buyer spanning several clouds and work systems may value an independent record and one set of rules. But it has to pay enough for that independence to cover the exceptions. Recurring outside evaluation paired with insurance is one possible way to price a bounded risk. **Layla:** Bounded is the key word. Insurance can make responsibility purchasable for a defined action with measurable losses. It cannot turn an unclear promise—“the agent will handle things”—into a contract anyone can price. I would look for narrow scopes, portable evidence, and real loss rates before calling this an enduring business. **Cato:** That updates our July read. Allocation gained support: routing can improve the cost-quality trade. But I put too much weight on settlement as the center of the story. Acceptance, verification, and recovery did more explanatory work over the full half-year. **Layla:** And a single model launch did not organize this shift. Even the controlled-access story around Mythos went quiet as a named thread. The practical question moved into procurement: what may the deployed system touch, what proof does it leave, and who can reverse it? **Cato:** Different scale now. Let’s leave the software operating layer and look at what it takes to make a promised data center run. **Cato:** My read is that booked demand is being mistaken for delivered capacity. Exelon’s high-probability data-center pipeline fell from roughly eighteen to eleven gigawatts after it required stronger collateral. The number changed when prospective customers had to put more at risk. That tells us something about the quality of the earlier pipeline, not simply its size. **Layla:** And stricter collateral could improve project selection. Some customers may have real plans but need more time to finance them. Others may have been holding a place they could not use. Which projects were delayed, canceled, or reassigned? The drop alone does not tell us. **Cato:** Agreed. Think of the stages separately. A customer asks for power. A developer secures land and equipment. Builders finish the site. The utility energizes it. The operator runs it, customers use it, and bills get paid. One stage can be valuable without proving the next has happened. **Layla:** Not energized yet. **Cato:** Exactly. SB Energy reportedly had a four hundred thirty billion dollars backlog with no data centers in service. That backlog describes promised future work. It does not describe servers already taking workloads. And I would not compare its dollars directly with Exelon’s gigawatts; those measure different things. **Layla:** A lender may still value the backlog. Long-cycle construction is normally financed before a site earns cash. Land, power access, and equipment may retain value if one customer changes course. The danger is not the existence of a promise. It is using the promise as though it were proof of operating service or eventual return. **Cato:** One pipeline shrank when buyers had to post stronger collateral; another is enormous before service begins—the shared mechanism is a financing commitment arriving earlier than reliable, revenue-earning capacity. **Layla:** And “earlier” can mean ordinary timing or a project that never converts. We need cohorts. Of the projects reserved in one period, how many get built, energized, utilized, and paid? How many are reassigned? That gives a better answer than either a backlog headline or one revision to a pipeline. **Cato:** It also matters to the utility. It may reserve time and engineering work for a project that later disappears. Collateral makes that claim on scarce planning capacity more costly to make casually. It can filter demand without telling us that every filtered project was worthless. **Layla:** Nor does it remove the local questions. A community can be asked to plan roads, water, or power upgrades long before the site produces the jobs and tax receipts in a developer’s pitch. The people asked to prepare for a project need to know which stage it has reached. **Cato:** So here is the narrow claim I will keep: a reservation is neither an operating service nor collected cash. I concede that neither Exelon’s revision nor SB Energy’s pre-service backlog establishes eventual returns. If these project cohorts convert into used capacity and paid bills at healthy rates, the gap we see now was normal construction timing. **Layla:** I agree with that boundary. If the projects stall, we will have learned something different: demand was being booked faster than power, financing, and construction could turn it into service. Both futures are still live. **Layla:** I suspect the leverage is moving to whoever can say, “I will depend on this result.” That would explain why review, operating controls, and collateral all became sharper tests—but they need not be the same business. A lawyer can reject a draft; a buyer can demand a recoverable agent action; a utility can require money behind a power request. If this read is right, we should see more accepted repeat work, fewer reopened failures, and more commissioned capacity earning cash. If those measures diverge, one conversion story was too tidy. **Cato:** Different receivers, different clocks. A software team can reopen a bug next week. A data-center project may take years to reach service. We should not force them onto one scorecard just because both began with a promise. **Cato:** Out of construction finance and into a shopping cart. The stakes are smaller, but the decision to trust an assistant is much more personal. **Cato:** My read is that daily activity proves curiosity, not durable delegation. Muse reached two hundred fifty thousand daily active users, yet usage reportedly faded against established shopping, phone, and calendar habits. One grocery error led to about one thousand dollars in unwanted food. That one mistake is not a general failure rate. It does show what a wrong action can cost a household. **Layla:** Start with the household, not the platform. People already have rules for who buys food, who checks the calendar, and which orders need a second look. An assistant enters those routines. The useful question is not “Will you trust it with everything?” It is “Which small action would you hand over again?” **Cato:** Suggest a cart, perhaps, but do not buy it. **Layla:** Yes. Or reorder the same coffee only below a price limit. Or draft a calendar change and ask before sending it to anyone else. Each permission has a different failure cost. That is why a count of prompts tells us less than repeat use by task and permission level. **Cato:** That’s the permission. **Layla:** And it can change after a mistake. If the assistant buys the wrong food, the household may still want meal ideas but revoke checkout access. From a product dashboard, the person remains active. From the household’s view, the assistant lost the right to act where money was involved. **Cato:** The strongest counter to my skepticism is that smaller, reversible actions could stick even if broad autonomy does not. Approval before purchase adds a step, but it may preserve the user’s willingness to try again. Over time, repeated correct suggestions might earn a wider scope. **Layla:** Or approval might become one more notification people ignore. We should measure whether the task actually saves time after corrections, whether users keep that permission enabled, and whether they repeat the action voluntarily. A retained user who only asks questions is not evidence that delegated shopping worked. **Cato:** So two futures remain. Broad delegation could stay fragile because a rare expensive error wipes out many small gains. Or narrow permissions could compound into a habit as people learn exactly where the assistant helps. Task-level retention and net value will separate them. **Layla:** And the household gets to set the boundary. That may be the most useful design fact here: trust is not a switch a company turns on. It is a sequence of specific permissions people can keep, narrow, or take back. **Cato:** At the start, I pictured a reservation on a developer’s spreadsheet. Now I picture someone arriving to use what was promised and deciding whether to come back. **Layla:** A reservation is a promise that something will be there when you need it. After all these handoffs, the harder question is who would rely on it a second time—and who makes it right if they can’t. **Cato:** Until next time. **Layla:** Take care, and keep asking.