On 23 September 2026, Amazon introduced a continuously running service called “workflows” for third-party sellers. Reuters confirmed that sellers can use a workflow to keep monitoring product prices or changes in ratings, and connect it through plug-ins to Amazon Quick and Anthropic Claude. The service is free and optional, and sellers can decide how much personal information to share. When Amazon released the agentic version of Seller Assistant in 2025, it had already said that the system could monitor inventory, account health and compliance, then act with a seller’s approval. The new development extends one-off questions and recommendations into tasks that continue to operate.
Two days earlier, Kevin Baum, Maximilian Kiener, Markus Langer and Johann Laux released the preprint “Anticipatory Human Oversight of Agentic AI”. They argue that conventional reactive oversight encounters a structural limit when an agent can decompose goals, call tools and act continuously over an extended period. Reviewing every action defeats the efficiency for which autonomous execution was introduced, while reviewing only aggregate results may come too late to detect a deviation produced by many locally reasonable actions. They propose supplementing real-time intervention with “anticipatory oversight”: specifying goals, values, constraints and escalation conditions in advance, then revising them in light of runtime records. This is a philosophical framework advanced in a preprint that has not yet been peer reviewed, not an empirically established consensus.
Together, these developments raise a more precise question than whether a human remains “in the loop”. If a person approves the initial task, can inspect the results and can stop the system at some point, is that sufficient to say that the agent is under effective human oversight?
Intervention must first be distinguished from oversight. Intervention is an event: approving an action, rejecting a request, pressing a stop control or taking over after an anomaly appears. Oversight is a relation sustained over time. A person’s purposes, reasons and role-based obligations must remain operative while the system selects actions, changes plans and handles exceptions. The presence of a stop control proves only that a person retains some causal power. It does not show that the person knows when stopping is required, or that the system continues to act in accordance with the person’s broader reasons when no stop condition is triggered.
Individual actions must also be distinguished from a trajectory of action. A seller may reasonably ask a system to reduce the price of slow-moving stock and may also reasonably ask it to protect product ratings. A persistent workflow may then repeat locally reasonable adjustments across different products, times and market conditions. Each action may fit its immediate instruction, while their cumulative result departs from the seller’s broader plans concerning brand position, inventory turnover, customer relationships or compliance. Local approval does not automatically amount to endorsement of the whole trajectory.
Philosophical accounts of agency commonly connect action with reasons, intention and control. What must be preserved here is not merely the human’s formal final authority, but a recognisable difference made by human reasons to system behaviour. If the seller’s long-term goals, tolerance for risk or compliance obligations change, the workflow should change accordingly. Nominal control without sufficient information, intelligible records, appropriate alerts and a realistic opportunity to correct the system cannot constitute substantive oversight.
My judgement is that human intervention is a condition of effective oversight, but it is not effective oversight itself. For a persistent agent, oversight must occupy at least three positions. Before operation, a person must define the scope of the task, prohibited actions, data boundaries and escalation conditions. During operation, the system must expose states relevant to those conditions and pause or request judgement when uncertainty or impact crosses a threshold. After operation, the person must be able to inspect the trajectory, identify cumulative departures and revise the constraints for the next iteration. If any of these positions is absent, a system can be formally assigned to a responsible human while leaving no one able to understand or govern the process in practice.
Sustenesis Theory makes the structure more explicit. Difference first separates the principal, the agent, the platform, an individual action and a trajectory across time. If these are merged into a single “system”, responsibility and control lose their point of attachment. Constraint does not mean limitation in general; it means the concrete conditions that make some actions possible and others unavailable, including permissions, tool scope, time boundaries, data access, thresholds, logs and escalation rules. Sustained Coherence is not the mere continuation of system operation. It is the structural consistency among human reasons, prior constraints, actual actions and subsequent correction, maintained through environmental change and feedback.
This analysis also marks the boundary between oversight and authorisation. Authorisation answers what a system may do. Oversight answers how human reasons continue to operate while the system is doing it. A workflow may hold legitimate permissions and still lack effective oversight. A closely supervised system may also be unable to act because its initial authorisation was illegitimate. The two relations matter together, but neither can substitute for the other.
Public material does not specify the maximum duration of Amazon workflows, their escalation thresholds, the granularity of audit records, error-recovery arrangements or reauthorisation mechanisms. Nor is there independent research measuring whether sellers can detect long-horizon departures. This essay therefore cannot determine whether the particular service has already achieved effective oversight. The present conclusion is limited to the conceptual standard: initial approval, possible intervention and retrospective inspection do not separately demonstrate meaningful human control over a persistent agent. Evidence is still required that human reasons, through constraints, feedback, records and correction, continue to shape the whole trajectory of action.
References
Amazon, “Amazon introduces agentic AI across the seller experience, transforming how sellers manage their businesses”, 17 September 2025, https://www.aboutamazon.com/news/innovation-at-amazon/seller-assistant-agentic-ai
Reuters, “Amazon rolls out new agentic AI for third-party sellers”, 23 September 2026, https://www.reuters.com/business/retail-consumer/amazon-rolls-out-new-agentic-ai-third-party-sellers-2026-09-23/
Baum, Kiener, Langer and Laux, “Anticipatory Human Oversight of Agentic AI: A Philosophical Account”, preprint, 21 September 2026, https://arxiv.org/abs/2609.24242
Stanford Encyclopedia of Philosophy, “Agency”, https://plato.stanford.edu/entries/agency/