TL;DR
- Supervising an autonomous AI agent is a different problem than monitoring a model that only predicts. An agent acts. Oversight has to control what it is allowed to do. Checking what it says is not enough.
- SR 26-2 excludes agentic AI from its scope by name. Federal model risk guidance is not the instrument examiners currently use to review an agent.
- The Conference of State Bank Supervisors released a new AI examiner framework on 16 September 2026, and it names agentic AI specifically. It is discretionary for each state, but it is the first US instrument built for this question.
- A real supervision program covers five things: inventory and ownership, decision boundaries, human checkpoints, full audit logging, and a working kill switch. Each one closes a different failure mode.
- Supervision breaks down in predictable places. Boundaries written down but not enforced, logs that capture outputs but not reasoning, human checkpoints that turn into rubber stamps.
- This piece is the Agentic Oversight cluster’s overview. The linked pieces on agent platforms, low-code agents, and enterprise accountability go deeper on each part.
- Caveat: the regulatory picture here is newer and less settled than model risk management. The CSBS framework is an early signal of what examiners will ask. It is not a fixed checklist yet.
A bank deploying an autonomous AI agent today runs into a specific gap. SR 26-2, the Federal Reserve’s replacement for SR 11-7, names generative and agentic AI in a footnote and excludes both from its scope, calling them “novel and rapidly evolving.” That does not mean an agent goes unexamined. On 16 September 2026, the Conference of State Bank Supervisors released an AI supervisory framework that state examiners can adopt at their own discretion, and it is the first US instrument to name agentic AI specifically. This piece defines what supervising an autonomous agent requires, maps that regulatory gap, and sets out the five components a real oversight program covers.
What supervising an autonomous AI agent means
A predictive model produces a score or a prediction, and a person decides what to do with it. An autonomous agent skips that step. It takes an action: it sends the outreach message, adjusts the credit line, flags the claim, executes the trade instruction, within boundaries someone set in advance.
That difference changes what oversight has to do. Monitoring a predictive model means checking whether its outputs stay accurate over time. Supervising an agent means controlling what it is allowed to do, watching what it did, and being able to stop it before the next action if something looks wrong. The word “supervision” fits better than “monitoring” here, because the job is closer to overseeing a person with delegated authority than to auditing a spreadsheet formula.
An institution that treats agent oversight as an extension of model monitoring, checking accuracy on a schedule and calling it done, is answering the wrong question. The right question is narrower and harder: what is this agent allowed to do on its own, and how would we know if it stepped outside that.
Why SR 26-2 does not cover your AI agents, and what is stepping into that gap
SR 26-2 and OCC Bulletin 2026-13, jointly issued 17 April 2026, replaced SR 11-7 as the guidance shaping US bank model risk management. The replacement states in a footnote that generative AI and agentic AI models are “novel and rapidly evolving,” and it places both outside the guidance’s current scope, with no date given for when separate guidance might follow. An institution cannot point to SR 26-2 and claim its agents are covered. They are named as excluded.
That gap is not going unnoticed at the state level. The Conference of State Bank Supervisors released its AI supervisory framework on 16 September 2026, built around a Core Examiner Guide, an Examiner Work Program, Nonbank AI Supplements, and an AI Use Case Risk Tiering Worksheet. The framework itself states plainly that it creates no new legal obligations. It is a discretionary tool, and each state financial regulatory agency decides how much of it to build into its own supervisory program.
Discretionary does not mean irrelevant. The framework is the first US supervisory document written specifically to give examiners a way to look at agentic systems, at a moment when the federal guidance most banks already know says nothing about them. An institution operating in a state that adopts this framework should expect its agents to be reviewed against it, whatever SR 26-2’s scope says.
None of this replaces ordinary model risk discipline for the predictive models an agent’s decisions are built on. Read Model Governance in Banking for how ownership and approval work for those underlying models, and AI Model Governance for what changes once a model generates content in addition to predicting it. An agent adds a layer on top of both. It has the authority to act on its own, something neither discipline alone covers.
What an examiner checks when reviewing an autonomous agent
Reporting on the CSBS framework describes examiners reviewing an agent along five lines: the boundaries on its actions, the human checkpoints built into its workflow, the logs it produces, whether its actions are reversible, and whether there is a working way to stop it. Each of these translates into a concrete practitioner question.
Boundaries on its actions. What is this agent authorized to do without a human approving first? Has anyone written that authorization down, or is it only assumed?
Human checkpoints. At which points does a person review the agent’s proposed action before it executes, and is that review a real decision point or a formality nobody has ever declined.
Logs. Does the record show why the agent acted, tied to the data and reasoning behind the decision, beyond only what it did?
Reversibility. If the agent’s action turns out to be wrong, can it be undone, and how long does that take.
Ability to stop it. Is there a tested way to halt the agent immediately, and has anyone tested it, or does it exist only as a line in a policy document?
An institution that can back up all five with evidence stands in a materially different position than one that only has a governance policy describing what should happen.
The five things a real agent-supervision program covers
Inventory and ownership. Every autonomous agent in production, named, with an accountable owner. This is the same starting discipline SR 26-2 does prescribe for models, applied to agents even though SR 26-2 itself does not name them.
Decision boundaries. A documented, enforced limit on what each agent can do without human approval. The limit has to live in the agent’s actual configuration. A policy statement about intended behavior alone is not enough.
Human checkpoints. Review points built into the agent’s workflow at the moments that matter, staffed by someone with the authority and the information to decline.
Full audit logging. A record that captures the reasoning behind each action along with the action itself, produced continuously as the agent runs, never reconstructed afterward.
A working kill switch. A tested, immediate way to halt an agent’s actions. Test it before it is needed. Do not assume it works.
An institution missing any one of these five has a partial program. An agent can have a named owner and no enforced boundary on what it does. It can have a kill switch nobody has tested since it was built. Each component covers a different way supervision fails, and none of them substitutes for the others.
Where agent supervision breaks down in practice
The gaps are predictable, and each one has a familiar shape. A boundary gets written into a policy document and never configured into the agent itself, so the agent can technically do more than anyone intended, covered in Agentic AI Platforms and the Control Gap. A low-code or business-user-built agent skips the review process built for engineering-owned systems entirely, because nobody classified it as an agent in the first place, covered in Low Code AI Agents in Regulated Workflows. An agent deployed by one business unit escapes the accountability structure another unit built for its own agents, because nobody owns the question of who is responsible when an agent crosses a boundary between teams, covered in Enterprise AI Agents: Where Accountability Sits.
The pattern across all three is the same one that shows up in traditional model governance. A control exists on paper. It was never tested against what the agent can do, who reviews its actions, or what happens when nobody planned for it.
How iTuring supervises every agent it runs on one platform
iTuring’s Agentic AI platform applies the same ownership, boundary, and audit discipline to autonomous agents that its Model Governance module applies to predictive and generative models. Every agent action is logged with the reasoning behind it. Maker-checker approval gates deployment and any change to an agent’s decision boundaries. Human-in-the-loop checkpoints sit at the points the institution defines, and a tested rollback path exists for every agent in production, including the ones a team might otherwise forget.
The same discipline underpins the 116% improvement in collections recovery iTuring delivered for a leading NBFC in India, where the outreach decisions driving that result ran inside boundaries the institution controlled from the start. Those boundaries were never an afterthought.
An autonomous agent that looks well governed on a policy document and one that can survive an examiner asking to see its boundaries, its logs, and its kill switch are not the same thing. Book a 15-minute discovery call to see how iTuring supervises every agent it runs, from inventory through a tested rollback path.
Sources
- “Supervisory Letter SR 26-2 on Revised Guidance on Model Risk Management,” Board of Governors of the Federal Reserve System, 17 April 2026. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm
- Conference of State Bank Supervisors, AI Supervisory Framework, released 16 September 2026. https://www.csbs.org/csbs-artificial-intelligence-supervisory-framework
- iTuring.ai case study, “Improve Collections and Optimize Efforts.” https://ituring.ai/case-study/improve-collections-and-optimize-efforts/
