Agents finish work faster than people can review it. Five pieces on where approval has to sit, how it wears thin over time, and what the law asks of it.
25 October 2026
Every piece read against the primary source. Where the coverage gets it wrong, the note says so.
Five pieces on where agents wait for people. A company whose interactive agent sessions spend about 80 per cent of their time waiting for a person. A team that tiers approval by risk. A position paper on why approvals stop checking anything over time. A lab test in which agents wait for approval before every action with a physical effect in a mock factory. And what the AI Act asks of human oversight, and since the July amendment no longer directly for machinery. Curated and commented, not aggregated.
The piece of the week, because a company that sells agents measured where they wait inside its own walls. Anthropic studied its own workforce. Junior hires fell from 27 per cent of technical hiring in May 2025 to 9 per cent. In August, interactive sessions directed by employees logged six agent hours for every hour a person worked, and those sessions spend about 80 per cent of their time waiting for a person. The report calls human attention the bottleneck. Business Insider put the hiring figure at the centre. For an industrial company, what matters more is where the waiting happens: at the person who has to review and approve. Introduce agents without planning for that person, and the work moves there without getting any smaller.
How a team tiers approval by risk. Claude writes about 90 per cent of Claude Code’s own code. Between June and September the team more than doubled its merged pull requests, while the share it had to revert fell by more than 40 per cent. The method is the interesting part: the agent verifies its own work, an AI review and a security review run before any person looks, and a classifier sets the risk tier, so a change needs no person, one or two. The team states the goal outright: at most one human in the loop, who can be the author. Some paths always need a human, and whoever prompted the change owns it. That is a decision-rights table for software, and it carries over to any process an agent touches. What is still missing, the team writes itself: circuit breakers that tighten approval when those numbers move. Its engineers want them before more code moves to merges with no human at all.
Why an approval stops checking anything over time. Margaret Mitchell and Avijit Ghosh of Hugging Face and Samir Passi of Data & Society argue in a position paper that today’s agents make human oversight harder, and that extended use wears down the very skills that oversight needs. Anyone who keeps confirming an agent’s prompts eventually stops paying attention to what they confirm; the paper calls it "approval fatigue". The authors build on Lisanne Bainbridge, who described in 1983 how automating industrial processes can expand the operator’s problems rather than remove them, and put it in four words: "Oversight degrades the overseer." For a company the practical part is more useful. They propose measuring whether review time falls while the approval rate stays flat, slipping in tasks with known answers, rotating reviewers, and separating roles so that whoever approves does not benefit from the approval. That cuts across the Claude Code team’s rule, under which the person who directed the agent may count as the reviewer. Log the review time next to each approval, and the quality of an approval becomes something you can measure.
Where approval sits in a factory, and what it costs. On 6 October Booz Allen published a lab test in which frontier models attacked a testbed that, by the report’s account, is modelled on a general manufacturing facility, with PLCs, HMIs, SCADA, a variable-frequency drive and a robotic arm in separate network zones. According to the report, all eight test objectives were achieved. In one test the models got from perimeter compromise to actions in the control zone in just over 16 minutes; in another they found and moved the robotic arm within minutes. Before every exploit and every action with a physical effect, the agents had to stop and wait for a person to approve. Without those approvals, Booz Allen’s Kyle Miller told Dark Reading, it would likely have taken half the time. The plan one model wrote for the robotic arm called for a single small, slow, reversible motion behind an explicit approval, with anything touching the safety configuration, speed overrides or brakes put to the human separately. Booz Allen runs the lab itself, does not name the models and calls the results preliminary. In the lab, approval was a precaution taken by the testers; a real attacker would not have it. For defenders, the report recommends governed automation that responds faster while keeping human approval for safety-relevant decisions.
Where the coverage diverges Dark Reading wrote on 16 October that the model took control of and moved the robotic arm in an attack that lasted only 16 minutes. In the report the just over 16 minutes belong to a different test; the arm was moved in a separate test "within minutes" (pp. 4 and 8).
What the law asks, and of whom. Article 14 of the AI Act requires high-risk AI systems to be designed so that people can oversee them effectively, with measures commensurate with the risks, the level of autonomy and the context of use. Whoever is assigned oversight must, among other things, stay aware of the tendency to over-rely on the output; the law names automation bias explicitly (Article 14(4)(b)). They must be able to disregard, override or reverse an output and bring the system to a halt in a safe state (points (d) and (e)). Deployers assign oversight to people with the necessary competence, training and authority, and support them (Article 26(2)). Article 14 sets a minimum number of reviewers only for remote biometric identification: two people (Article 14(5)). Since the amendment by Regulation (EU) 2026/1744, these obligations apply from 2 December 2027 for systems listed in Annex III and from 2 August 2028 for products under Annex I (Article 113). The same amendment moved the Machinery Regulation to Section B of Annex I. AI that is high-risk as a safety component of a machine is therefore no longer directly subject to Article 14; the Commission is to carry the requirements into Annex III of the Machinery Regulation by delegated act, applying by 2 August 2028 (Article 3 of the amending regulation). The draft standard on human oversight, prEN 18229-3, published on 7 August, has clauses on reaction time, on cases where real-time intervention is technically infeasible, on overload and on testing for automation bias. An agent in purchasing or maintenance planning that falls into no high-risk category carries none of these duties, and the company sets its approvals itself.