The orchestration layer
An AI agent can now hand work to hundreds of helper agents in minutes. No one can approve them one by one.
In July I told the Financial Stability Board to govern the agent in charge instead, the one that hands out the work (the orchestration layer). On October 15 the FSB published its final sound practices for AI in financial services, and some of the critical controls I argued for are now in it. (The practices are guidance for boards and senior management to reference, not rules.) Side by side:
What I argued: oversee at the orchestration boundary, not agent by agent.
The final report: “oversee agents and sub-agents at the orchestration layer where possible.”
What I argued: treat spawning a sub-agent as a delegation, and govern it.
The final report: “impose appropriate, consistent deterministic controls governing delegation to sub-agents.”
Partly: I argued that where the check is a limit, it should compute rather than be another AI that reasons.
The final report, more broadly: higher-risk agent use cases “may warrant layered deterministic controls.”
None of those lines were in the June draft.
About 160 responses were filed, and several asked for more on agentic AI. I can’t know whose shaped the text.
Three controls the final report does not include, but I believe are essential when agents create other agents:
- Limits that apply to the whole tree of agents, so splitting a task across sub-agents cannot slip under a threshold.
- By default, a sub-agent gets no more access than the agent that created it.
- The agent that approves sits outside the tree it approves.
The FSB’s own overview calls its agentic practices still emerging. For higher-risk agents, these three are where I would start.
- Financial Stability Board, 15 October 2026: Sound practices for the responsible adoption of artificial intelligence, final report. All quoted lines are verbatim from the final report; the response count and the “still emerging” description are from its overview.
- The June 2026 consultation draft, checked for each quoted phrase. None of the quoted lines appear in it.
- My comment and supplemental, as filed: Bidun Group (PDF), filed 15 and 19 July 2026.
Scope note. Around 160 responses were filed and several asked the FSB to expand its treatment of agentic AI. No claim is made here that the FSB adopted any language because of this comment; the side-by-side states presence, not cause. The sound practices are guidance for boards and senior management to reference, not rules.
Written for compliance and risk readers in regulated financial firms. Informational, not legal advice; Mike Bidun is not a lawyer.