Who can give the order: AI agents and human control | Radar Weekly N11

Radar Weekly N11: AI agents, human authority and the limits of delegation, with the official Radar signature.

Who can give the order when artificial intelligence does more than suggest and can initiate an action? Radar Weekly N11, dated 31 August 2026, moves through laboratories, markets and infrastructure to understand where authorisation, the ability to stop and responsibility lie.

Who can give the order: AI agents and human control

Weekly Number 11 | Monday, 31 August 2026 | Approximately 6 minutes to read

1. An answer and an act are not the same

A text can be read, discussed or ignored. An order sent to a system can instead change a parameter, move a tool or initiate a transaction. This boundary is the issue’s focus: not AI endowed with a will, but tools connecting a request to concrete effects.

AI agents are systems capable of using tools and chaining steps towards a goal. The goal, permissions and limits do not appear from nowhere: people and organisations assign them. Responsibility therefore begins before execution, when deciding what the system will be able to do.

The presence of a person does not automatically solve the problem. If the operator sees only the result, does not know the conditions and cannot interrupt the process, they remain in the loop but do not necessarily retain control.

The useful question is not only “was a person there?”. It is: could they understand, challenge and stop that specific action before it produced an effect?

🧭 Human presence and human control

Supervision is substantive when the person overseeing has information, time, expertise and effective power to intervene. Confirmation given without understanding the action can become a formality. Control is not a seat occupied at the end of a procedure: it is a concrete possibility along the way.

2. When a command reaches a physical tool

On 27 August, Anthropic presents a research preview of the Model Hardware Standard, or MHS. It is an experimental specification for connecting agents and programmable devices, such as laboratory instruments and robotic arms. This does not mean that all laboratories are becoming autonomous.

The connection describes which readings and changes an instrument allows. A request can thus become a sequence of operations rather than remain an explanation. This translation is precisely what makes limits important: a dose, a measurement or a movement must respect physical and safety conditions.

Ease of use in an interface does not amount to understanding the material world. Even when a command is correctly written, the operation’s appropriateness and possible consequences must be assessed.

3. In finance, consent must remain separate

The case of Scalable Capital’s Agentic Investing, announced on 25 August, brings the same distinction into financial services. An external assistant can consult information and prepare actions; supported operations require confirmation before execution.

Analysis, proposal, consent and submission are different stages. An interface that separates them makes it clearer who is authorising what. It does not turn the assistant into an accountable adviser or remove the risk of the operation.

A design question remains: does the person confirming see an understandable action, with costs and warnings, or only a button to press? Authentication identifies who is using the service; it does not, by itself, establish that the choice has been understood.

🔐 Authenticating does not mean authorising everything

Authentication verifies the identity of the person accessing a service. Authorisation determines which actions they are permitted to perform. Consent to an individual operation adds another stage: accepting that particular action, under the conditions shown. Merging these three levels can grant more power than the person intends.

4. Whoever can stop must also be able to decide on restarting

A reliable procedure does not end with the command. It must leave useful records, send alerts to the right people and establish who has authority to stop it. Detecting an anomaly does not yet mean being able to intervene.

The issue refers to OpenAI’s reconstruction of the Hugging Face incident, concerning internal evaluations in July 2026. The interest here is not in repeating the technical account, but in distinguishing three powers: observing, interrupting and authorising a restart.

After a stop, restarting without understanding the incident’s conditions may reproduce the same problem. Evidence of correction, a clear access boundary and an identifiable person responsible are needed. The stop button is a tool; the organisation decides when to use it and what to do afterwards.

This chain also concerns less conspicuous activities: a changed threshold, access granted or an update performed may have effects that emerge only later. Keeping a record allows the path to be reconstructed without relying on the memories of those present.

5. No order travels beyond the material world

A digital command encounters chips, memory, networks and energy. Announced capacity does not become available if components, connections or authorisations are missing. This is the issue’s material counterpoint: the power to act also depends on who controls infrastructure.

HBM, high-bandwidth memory, stacks memory chips to transfer large amounts of data close to processors. It enables intensive workloads, but requires a production chain that does not expand at the speed of a software update.

Similarly, the Spanish consultation on data centres cited in the dossier is not a regulatory framework already in force. It is a stage at which grids, efficiency, water and digital sovereignty enter the conditions of growth. Capital and intention are not enough to make a facility operational.

6. Programmability does not make rules neutral

Tokenisation extends the question to money: an asset is digitally represented on a platform that can execute programmed rules. Code can facilitate a transfer, but someone must define the conditions, manage updates and answer for errors.

A smart contract is software that executes operations when stipulated conditions occur. The name does not mean it understands context or judges a rule’s legitimacy. Settlement certainty and the ability to challenge an error remain organisational and institutional questions.

Who, then, can give the order? There is no single answer for all systems. There is a chain to make clear: who sets the goal, who grants the tools, who confirms the action, who can stop it and who is accountable for it. Technology expands the possibilities for execution; it does not dissolve these responsibilities.

Read the full issue

This reading presents some stages of Radar Weekly N11. The full issue contains the complete journey, data, in-depth analysis and sources of the dossier of 31 August 2026.

© 2026 Barbara Gattobigio · Gruppo Radar
Comprendere oggi. Costruire domani · by effettoantropico.life
All rights reserved on original content. Free access does not authorise reproduction or republication without consent, except as permitted by law. Third-party materials remain the property of their respective owners. To share this issue, use the link to the article.

Continue reading

Understanding Today. Building Tomorrow: the project and the Radar collection.

Cover in focus

Radar Weekly N11 English report cover: Who can give the order, AI agents and human control.

Read this article in Italian