The person of record

Working model. An engineer may use agents to produce calculations, drawings, comparisons, and evidence, yet still have to sign the result. The interesting constraint is not whether the agent can perform those visible tasks. It is why the final decision remains attached to that engineer.

Three questions help separate the reasons:

1. Can the agent obtain the context needed to act competently?
2. Can someone check what it did and what happened as a result?
3. Can the right to commit, attest, decide, or accept liability move away from the named person?

An agent may pass the first two tests and still fail the third. It may also have permission in principle but lack access to the factory, patient, negotiation, or local exception that makes competent action possible. A capability demo usually answers only a small part of this.

Why work remains attached to a person

Sometimes the missing context has not been captured. A technician notices a sound, a clinician sees a change, or a long-serving operator knows which exception the manual omits. Better sensors, models, and documentation can move some of this knowledge into machine-usable form. I would not treat current tacit knowledge as permanent protection for an occupation.

Sometimes the context is digital but must stay in one place. Patient data, classified systems, factory controls, or a deal room may not be available to a central service. This limits distribution more than automation: a model can be brought into the hospital, factory, or secure environment.

Relationships create a different constraint. A record can describe a client’s history or a diplomat’s contacts, but it does not automatically transfer the other party’s willingness to trust someone. Agents may prepare and operate the surrounding workflow while a particular adviser, physician, or negotiator remains part of the productive relationship.

Institutions also bind decisions to people deliberately. An engineer, physician, auditor, director, fiduciary, judge, or responsible officer may have to answer for an outcome even when the underlying analysis is heavily automated. That is an accountability choice, not missing data.

These constraints can change. Trust may move from a person to an institution or certified system. Law may move liability from a professional to an employer or vendor. Those changes are political and cultural as well as technical, so a model benchmark cannot predict their timing by itself.

Human-anchored work may still be mostly automated

The engineer of record does not need to draw every line to remain accountable for the design. The role may move toward setting acceptance criteria, challenging assumptions, resolving exceptions, approving material claims, and accepting residual risk. A senior analyst may retain ownership of definitions, disputed adjustments, and attestations while agents perform most transformation and reconciliation.

This changes the capacity question. Instead of asking only how many people a team employs, ask how much agent-operated work one accountable person can genuinely review. The answer depends on consequence severity, exception rate, evidence quality, volatility, and the amount of relationship work involved.

The dangerous version is easy to recognize: one person signs for hundreds of processes but cannot tell which assumptions changed, which cases were excluded, or what evidence supports the result. The name remains on the form while the ability to exercise judgment has gone elsewhere.

Protecting against that requires more than a signature. The system needs to escalate material exceptions, show evidence in a form the person can challenge, limit delegated scope, detect overload, allow refusal, and record what was never presented for review.

Management may divide along similar lines. Agents can carry status, assign routine work, chase dependencies, and translate updates. Resolving incompatible preferences, maintaining trust, allocating legitimate authority, and deciding which ambiguity matters are likely to remain attached to people longer.

I am not confident that every profession will retain a permanent human core. The narrower claim is that task automation does not settle the question. Where local judgment, a trusted relationship, or accountable approval must remain bound to a named person, the surrounding workflow can become agent-operated without making that person ornamental. Evidence that institutions can transfer those three functions safely would weaken this model.

Use this note as a worked model, not a forecast. The transferable question: when most of a professional workflow can be performed by an agent, what is actually holding the final decision to a named person, and would that thing survive a change in law, trust, or instrumentation? The worked model separates three reasons and insists they be tested apart: whether the agent can obtain the context needed to act competently, whether someone can check what it did and what followed, and whether the right to commit, attest, decide, or accept liability can move away from the person at all. An agent may pass the first two and still fail the third; it may hold permission in principle while lacking the access that makes competent action possible. Tacit knowledge is treated as movable rather than permanent protection, while relationships and deliberate institutional accountability are treated as different constraints with political rather than technical timelines. The capacity question becomes how much agent-operated work one accountable person can genuinely review, and the failure mode is a name on a form with the judgment gone elsewhere. Apply the question to a role you understand. Test the three separately, and name what the system would have to escalate, evidence, limit, or refuse for the accountable person to stay real. Say where your constraints diverge -- no licensure, transferable trust, an employer already carrying the liability. Produce a per-reason assessment with the evidence that would weaken it.