Breadth Got Cheaper. Depth Did Not.

Entering a specialized field has historically been expensive, and becoming an expert is harder still. A software engineer has spent years learning how systems fail. A machine-learning engineer understands experimental design, model behavior, and the many ways data can mislead, while infrastructure, databases, interface design, and academic research each have their own language, tools, and accumulated judgment.

Before an outsider could do useful work in one of those fields, they had to learn its vocabulary, install its tools, find its canonical sources, reproduce its standard techniques, and become familiar enough with the work to ask useful questions. The first serious attempt might require weeks of preparation, and even when a specialist was available, the outsider often lacked the context needed to evaluate the answer. Agents reduce much of that entry cost, but they do much less to compress the years of consequences that produce expert judgment.

That creates an uncomfortable split. Expertise remains difficult to achieve, but access to a profession’s standard methods is becoming less scarce. Agents make the conventional middle of many fields more accessible through documented methods, familiar patterns, standard implementations, textbook explanations, and known failure modes. A field can therefore lose some of its market power even while its deepest practitioners remain difficult to replace.

In AI Has Not Removed the Learning Gap, I argued that agents compress implementation knowledge much more than the judgment required to make software hold together over time. That has a second consequence: if baseline competence arrives faster while depth remains difficult, one person can cross into more fields without becoming an expert in all of them. AI has made breadth cheaper, not depth.

The Middle Is Larger Than We Admit

Professional identity tends to emphasize the hardest parts of a field. Engineers talk about architecture under change, not wiring another API. Machine-learning researchers talk about novel methods and subtle experimental flaws, not fitting a standard classifier. Infrastructure specialists talk about distributed failure, not deploying a conventional service, while designers talk about understanding human behavior, not arranging a competent form.

Those distinctions are real, but they obscure how much professional work takes place below the expert edge. Experts are right that agent-assisted work is often not expert work. Professional status does not guarantee expert output either, because every field already produces conventional, uneven, and mediocre work alongside its best work. Agents do not need to outperform the best practitioner to enter a field; they need to make its standard methods accessible enough to compete with the work the field already accepts.

Most tasks do not require a novel algorithm, an unusual distributed protocol, or a new interaction model. They require someone to recognize the class of problem, select a reasonable method, apply it correctly, and check the result. That is not trivial work, but it is patterned work with terminology, examples, tools, and review criteria that an agent can help retrieve and apply. In an experiment with 453 professionals, access to ChatGPT made common professional writing tasks faster, improved their rated quality, and reduced the performance gap between participants. Agents compress the distance to that kind of competence.

A software engineer can ask an agent to explain feature selection, construct a baseline model, identify likely leakage, and produce an evaluation pipeline. The result will not reflect the judgment of an experienced machine-learning practitioner. It may still be good enough to test whether an idea deserves further attention.

The exchange works in the other direction. A machine-learning engineer can use an agent to package a model, build a distributed prototype, add baseline observability, configure conventional infrastructure, and reason through common failure scenarios. The resulting system may not impress an infrastructure specialist. It may still work and teach its creator something important.

The same pattern extends across databases, interface design, operations, analytics, presentations, and academic work. A database outsider can compare indexing strategies and inspect a query plan. A backend engineer can prototype a usable interface rather than stopping at an endpoint, and a practitioner can find the established literature around a question, map competing explanations, and reproduce a conventional analysis. None of these people have crossed all the way into expertise, but they have crossed far enough to participate.

Useful Is Not The Same As Expert

Crossing a professional boundary does not remove the distinction between useful work and expert judgment. Agents are unusually good at producing the appearance of finished work: a model trains, a service deploys, a query gets faster, a screen looks polished, and a literature review has citations. Each result can be locally convincing while remaining wrong in the ways that define the deeper craft.

The machine-learning pipeline may encode the target in a feature. The distributed service may work until retries amplify an outage. The optimized query may trade one fast path for unpredictable write performance. The interface may be visually coherent while making the user’s actual task harder. The literature review may summarize papers accurately while missing the disagreement that matters.

Baseline competence applies known methods to recognizable problems. Expert judgment determines whether the problem has been recognized correctly and whether the method’s assumptions hold. An agent can describe common risks and build a checklist from them, but it cannot reliably know which risk dominates when the decisive context is tacit, political, missing from the prompt, or not yet visible.

How far an outsider can safely cross therefore depends on verifiability. Agent assistance travels farther when a result can be checked against a test, benchmark, reproducible calculation, or fast real-world response. That unevenness appeared in a field experiment with 758 consultants: assistance improved speed and quality on tasks within the model’s capabilities, but made participants less likely to reach the correct answer on a task outside them. When feedback is delayed, evidence is ambiguous, or mistakes are expensive, the expert’s ability to recognize the missing question remains the bottleneck.

Crossing into a field gets you to its standard methods faster. It does not give you the history of consequences that tells an expert when those methods no longer fit.

The Symmetry Is Uncomfortable

People readily notice weak agent-assisted work in their own field. An engineer sees generated code and finds the abstraction careless. A designer sees a generated interface and finds the hierarchy obvious but empty, while a researcher sees an analysis and finds the inference unjustified. The expert concludes that AI cannot do the work.

Often, the criticism is correct, but the conclusion is too broad. The same expert may be using an agent to perform similarly conventional work in another field, with weaknesses that are equally obvious to its practitioners. The engineer’s statistical analysis and the researcher’s deployment architecture may both be amateurish by expert standards. They may also both be useful.

The relevant comparison is not always between an assisted outsider and the best specialist. It is between what the outsider can attempt now and what they could reasonably attempt before. A competent engineer who can test a conventional ML approach has gained useful range even if they would not publish the method, just as an ML engineer who can operate a distributed prototype has gained useful range even if they should not design a bank’s transaction platform. The capability is meaningful without being complete.

No profession is exempt from this symmetry. If an agent can help you cross into someone else’s field, it can help them cross into yours. Cheaper participation also gives outsiders enough vocabulary and context to interrogate specialist work.

Expertise Becomes Easier To Question

Agents change more than who can produce work. They change who can interrogate it. An outsider can ask for an explanation of the method, translate specialist language, compare a proposal with conventional practice, surface disputed assumptions, and generate questions for review. This does not make the agent a truth oracle; it makes the agent a translator and review scaffold.

That difference is important. Asking an agent whether an expert is correct may only replace one unexamined authority with another. Asking it to explain the expert’s assumptions, identify the evidence required, present plausible alternatives, and show where specialists might disagree creates a better basis for judgment.

The outsider still may not be able to resolve the disagreement. But they can often move from “I do not understand this field” to “I understand what this decision depends on.” That is a substantial change in the relationship between specialists and everyone around them. Expertise still deserves weight, but easier evaluation may make that weight less automatic.

Titles Prove Less

Many professional moats have combined genuine judgment with costly signals: credentials, job titles, specialized vocabulary, prestigious institutions, and years spent near the work. Those signals remain informative, but as standard methods become easier to retrieve and execute, they become less conclusive on their own.

A title tells us where someone has spent time, not how well they can carry an unfamiliar problem. A textbook method shows that someone knows the standard approach, not that they can detect when it should be abandoned. Years in a field can produce judgment, but years alone do not demonstrate it.

Academia is not exempt from this distinction. Credentials and years in a discipline do not make every application of its methods expert. An agent can explain a standard method, apply a textbook technique, inspect an experimental design, or help an outsider question an interpretation. That makes the method less exclusive. It does not supply a consequential research question or guarantee that the evidence supports the claim.

What distinguishes depth is increasingly visible in choices:

• Did you frame the right problem?
• Did you notice which assumption would break?
• Did you preserve the important boundary?
• Did you distinguish a persuasive result from a reliable one?
• Did you know when the conventional method was insufficient?
• Did you involve a specialist before the cost of error became unacceptable?

This is where depth becomes more legible. Expertise is not possession of the method. It is judgment about the method.

The Entry Level Carries The Cost

Cheaper access is an opportunity for the person crossing into a field, but it is a threat to the person trying to enter that field as a profession. Junior work often consists of the standard methods agents make cheaper: routine integrations, baseline analyses, test coverage, documentation, first drafts, and well-understood changes. Companies paid junior professionals to do that work partly because senior people had more consequential problems to solve. The arrangement also gave juniors the repetitions, mistakes, and exposure they needed to develop judgment.

Agents can now perform or accelerate much of that work. If a smaller group of experienced people can direct agents through the routine workload, a company has less immediate reason to hire someone who still needs supervision. Entering a field technically becomes easier while entering it professionally becomes harder.

Junior workers contribute observation, context, coordination, and developing judgment beyond their routine output. Routine output has still supplied much of the economic justification for hiring and training them, and that justification is weakening before the need to develop future experts has disappeared. A company can reduce its entry-level roles, improve short-term throughput, and still create a long-term succession problem in which expertise becomes more important while fewer people are paid to develop it. Addressing that problem requires treating skill development as a separate responsibility rather than assuming that more generated output will create future judgment.

Breadth Lets Depth Travel

The practical response is not to abandon specialization. It is to stop treating the edge of a specialization as the edge of responsibility.

An engineer can carry an idea through data analysis, interface design, deployment, and explanation before asking for help. A researcher can build the software required to expose a method to real users. A designer can investigate implementation constraints directly. A database specialist can prototype the product behavior that makes a storage decision relevant.

That breadth includes work technical specialists often treat as someone else’s responsibility. The same person can investigate a market, model a scenario, prototype an interface, analyze usage data, prepare a proposal, build a presentation, and explain why the result matters. None of those outputs must be expert-level to move an idea forward. Competent work across the whole path can be more valuable than isolated expertise at one step.

Agents also make the communication layer easier to practice. They can turn a pile of notes into several possible arguments, simulate a skeptical audience, expose unsupported claims, and help separate features from outcomes. They cannot decide what is true, what matters, or what the author genuinely believes. That responsibility does not become cheaper.

The goal is not to remove experts from the process but to make handoffs more deliberate. Specialists are most valuable where assumptions are unclear, feedback is delayed, failure is expensive, or the system must survive conditions that a standard implementation does not capture. They are less necessary as gatekeepers to every routine technique in their field.

That feels like erosion to professions accustomed to owning both the methods and the judgment. It is better understood as a separation between them: agents make methods easier to access, but they do much less to commoditize responsibility for the result. The durable advantage is therefore neither narrow knowledge protected by a difficult entry path nor shallow familiarity with everything. It is depth in at least one place, enough breadth to connect that depth to an outcome, and the judgment to know where competence ends.

Agents flatten the middle and widen the territory one person can explore, but they do not flatten the edge. That edge is simply harder to claim by title alone because it has to show up in the decisions.