Regulation

The FSB's final AI sound practices add agentic AI considerations: minimum permissions, kill switches and human-in-command oversight

The Financial Stability Board published its final report on Sound Practices for the Responsible Adoption of AI on 15 October 2026. It sets out 12 practices, says the report now includes additional considerations for agentic AI and multi-agent systems, and states there is little evidence at present of GenAI being used for fully autonomous trading.

· Source: Financial Stability Board


On 15 October 2026 the Financial Stability Board (FSB) published the final version of its Sound Practices for Responsible Adoption of Artificial Intelligence (AI), along with an overview of the responses to the June consultation. The press release describes 12 sound practices in two groups: organisation-wide governance (strategy and oversight, roles and accountability, AI risks in the risk framework with documentation, and organisational adaptability) and a risk framework for the AI lifecycle (materiality and risk assessment, selection, data governance, explainability, performance management, human oversight, cyber and ICT risk, and third-party AI risk). The FSB strongly encourages boards and senior management of financial institutions to reference them, and calls them a practical toolkit that firms can tailor to their context and risk profile.

The change that matters most for readers who build or run AI trading agents is how the final report treats agentic AI. It says the report now includes additional considerations for managing advanced risks associated with agentic AI and multi-agent systems, while stating that practices specific to agentic and multi-agent systems remain nascent and may warrant more comprehensive treatment in the future. The report lists agentic risks that can materialise at great speed: unauthorised actions, erroneous actions that may only appear once an agent is live, data exposure, disruption to connected systems through APIs and other agents, the impracticality of real-time human monitoring as use scales, memory poisoning, and weakened governance when an agent spawns sub-agents. On access, it says the risks of an agent's tool and system access can compound multiplicatively rather than additively as permissions expand, so firms should weigh new access together with existing access and agree a risk tolerance for each agentic use case at inception.

On oversight, the report describes a human-in-command model, in which humans decide the extent of autonomy, set guardrails and manage overall impact, for AI with high levels of autonomy such as agentic AI. It also describes a kill switch as a mechanism enabling human intervention to halt or constrain AI operations, from complete shutdown (which it notes can be used in disabling algorithmic trading) to graduated degradation where systems move from autonomous to human operation. One illustrative case study, an insurance company rather than a trading firm, describes agents given the minimum permissions needed, individually registered and monitored, with logs of intermediate steps, periodic scheduled stops, and random pauses to confirm that sub-agent activity stops when the parent agent does. The report says its case studies are illustrative only and do not prescribe an approach.

What this does and does not mean for trading. The FSB's report is not an international standard and does not bind any firm or individual user; national supervisors decide whether to draw on it. On trading specifically, the report says there is little evidence at present of GenAI being used for fully autonomous trading, while financial institutions are exploring agentic AI that can sequence trading-related tasks such as scanning news, generating signals and preparing execution instructions. That is a description of the institutional sector, not guidance for retail users of trading agents, and nothing in the release or report is investment advice. Our 13 October briefing covered the Chair's letter and could not read the report itself; this one reads the published final report and the press release. What to watch next: whether the UK, EU and US authorities cite these practices in their own expectations, and the FSB's own review as agentic practices mature.


Read the original: Financial Stability Board - The FSB's final AI sound practices add agentic AI considerations: minimum permissions, kill switches and human-in-command oversight. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.