Every conversation about delegating work to agents eventually hits a question that has no agent-shaped answer: when this is wrong, who is answerable? You can hand over the writing, the running, even much of the deciding. You cannot hand over being the one who answers for the result. That residue — the part that does not delegate — is the subject here. We call it the accountable core.
In the three-layer model — tasks delegate to agents, the work reconstitutes a layer up as the roles a human holds — there is a floor beneath the roles. Below the decomposer, the reviewer, the arbiter and the owner sits something none of them, and no agent, can absorb: answerability. Naming it matters, because it is the thing that does not scale with the number of agents you point at a problem, and it is increasingly the only thing a senior engineer is uniquely there to provide.
What the accountable core is — and is not
It is not authorship: the agent wrote the code, and that changes nothing about who is answerable for it. It is not execution: the pipeline ran it. It is not even judgement in the abstract — AI can widen who is capable of making a good call. The accountable core is narrower and harder than any of those: it is being the named human who owns the outcome and its consequences, the person whose name is under the decision when it is examined later.
Why “the AI did it” is never an answer
Picture the three audiences that arrive after something goes wrong: the customer whose payment silently failed, the regulator asking who approved a change, the incident review asking why a system behaved as it did. To none of them has “the model generated it” ever been an acceptable answer. Accountability needs a person to attach to. The academic literature circles this as the “responsibility gap” — work whose authorship is a machine and whose accountability no human has actually claimed. The gap is not a technical defect to be patched; it is a hole that only a named owner fills.
Judgement can be widened. Accountability cannot.
It is tempting to comfort ourselves that judgement is the safe human preserve. The honest economic reading is sharper. Autor argues that AI can democratise expert judgement — push high-stakes decision-making down to more people than could exercise it before. So “who may judge” is genuinely widening. But that is a different axis from “who is answerable.” AI can help more people make a sound decision; it cannot be the one held to account for it. The accountable core survives precisely because it is not about capability — it is about answerability, and answerability has to land on someone.
Operationalising it
The accountable core fails quietly when it is left implicit — “the team” owns it, which means no one does. Make it concrete:
- A named owner per change — a person, not a team or a queue. Accountability does not distribute.
- An explicit acceptance signature — the moment a human takes ownership of an agent-produced change is recorded, not assumed.
- A reconstructable trail — who decided what, when, with which context (provenance), so the core can be examined after the fact.
- The core made visible in the lifecycle — an acceptance gate where ownership is transferred, rather than a silent merge.
None of this slows a capable team down; it is mostly making explicit a thing good engineers already do. What it prevents is the failure mode of abundance — shipping unowned decisions at machine speed because the question of who answers for them was never asked.
As agents take more of the tasks, the work that remains shrinks in volume and grows in consequence. The accountable core is the part that does not get cheaper as generation does. It is where senior value concentrates, and it is the floor under everything else: you can delegate the work; you cannot delegate being the one who is answerable for it.