Not long ago, working with AI meant watching it work. You submitted a prompt, it replied in seconds, and every step was in front of you. If it went wrong, you saw it happen. That era is ending.
Self-directed agents — AI systems that accept a brief goal and then run, without interruption, for hours or days at a stretch — change the accountability relationship in a way that most organisations have not yet come to terms with. The output arrives finished. Hundreds of decisions happened inside it. None of them were witnessed.
The new shape of human work
The model that survives this shift has two poles and a missing middle. At the front: a short specification — what you want, roughly why, and any constraints that genuinely matter. At the back: a judgment call on what came back. The long middle stretch — tracking progress, catching drift, redirecting, coordinating loose ends — is what autonomous agents absorb. For organisations that relied on a management layer to hold that middle together, the structural implication is significant.
The human job is now specification and judgment, not supervision. Both halves demand more craft than they used to, because the agent’s output is no longer a draft that shows its working — it is a finished artefact that embeds its working silently.
What “embedded decisions” means in practice
When a capable agent completes a multi-step task, the deliverable it returns encodes dozens of choices you never saw made: which sources it weighted, which edge cases it resolved one way rather than another, which ambiguities it decided to ignore. The agent did not surface them. It resolved them and moved on.
This is not a flaw — it is precisely the capability that makes autonomous work useful. But it creates a specific accountability problem: the person who set the goal is answerable for the output, including the choices embedded inside it.
Agency law — the legal framework governing situations where one party acts on behalf of another — already handles this for humans. An employee who signs a contract on your behalf, or a lawyer who settles a dispute in your name, binds you to their decisions even if you were not in the room. The same doctrine now applies to machine agents acting in your name. What the agent commits to, agrees to, or produces is your commitment, agreement, or product. Organisations that are still treating AI agent output as a first draft to skim are misreading the liability exposure.
The last two percent is the whole problem
Autonomous agents are not yet fully reliable, and near-complete reliability is a different problem from low reliability. An agent that fails thirty percent of the time is easy to manage — you build human review into every step and treat it as a drafting assistant. An agent that succeeds ninety-eight percent of the time is harder, because the apparent track record encourages skipping the review, and the two percent of failures are the ones most likely to involve the highest-stakes decisions: the cases the agent found ambiguous, where it made a judgment call rather than following a clear path.
The enterprise value of autonomous AI sits in the gap between ninety-eight percent and one hundred percent trust. Closing that gap is partly a model capability question and partly a review practice question — and only one of those is within the practitioner’s control today.
What good back-end judgment looks like
Checking work you never watched get made is a different skill from editing a draft you watched take shape. A few principles hold up across contexts:
Reconstruct the decision surface. Before evaluating the output on its own terms, ask what choices the agent must have made to produce it. Where the task was genuinely ambiguous, those are the places worth examining — not to second-guess everything, but to confirm that the implied choices are ones you would have endorsed.
Test the edges, not the centre. The main body of an agent’s output tends to be where it is most confident, and also where it would have been fine either way. The value of review sits at the boundaries: the places where the task got complicated, where competing priorities existed, where a different reasonable choice would have changed the output materially.
Make the specification sharper next time. The gaps that the agent filled silently are usually the same gaps that the specification left open. A review is also a retrospective on the brief.
The practice is still young and the tooling to support it is uneven. What is already clear is that reviewing finished autonomous work is a first-class professional skill — not a formality, and not something that can be treated as optional once the output looks plausible on the surface.
Text summarized and optimized using Anthropic’s models and reviewed by a human.