The Citibank payment error happened even though the procedure, warning, and three approvers all appeared to work correctly.
2
AI agents need autonomy with boundaries, visibility, and runtime controls instead of total sandboxing or unrestricted access.
3
Five composable controls are an agent registry, risk-based review, purpose-based tool access, shared evals, and classifiers that reduce approval fatigue.
Summary
John Feminella uses Citibank's 2020 payment error to explain why AI agent safety cannot depend on a single warning, approval step, or policy document. Citibank intended to send about $8 million in interest while keeping loan principal inside the bank. A processing mistake sent about $900 million to more than 300 lenders. The warning was accurate, and three approvers accepted the transaction because they understood the payment as legitimate. Feminella argues that agents should receive bounded autonomy, much as human workers do. He proposes five controls: register every agent and its ownership, tools, and endpoints; review agents according to their risk; let declared purpose determine tool access; provide shared evaluation infrastructure; and use classifiers to route only ambiguous or dangerous actions to people. Together, these controls allow agents to operate without forcing every experiment through the same slow process or granting every agent broad permissions.
Controls can work as designed and still allow a catastrophic mistake
Feminella begins with Citibank's August 2020 payment error. Revlon owed about $8 million in accrued interest, while the principal on its loan was supposed to remain in place. Citibank used Flexcube to model the transaction as if the loan were being paid off, then intended to route the principal to an internal wash account. The required boxes were not checked. Flexcube displayed a warning that money was about to leave the bank, but that warning was also consistent with the intended interest payment. Three approvers accepted it. The system sent about $900 million to more than 300 lenders.
Agent safety should allow bounded autonomy rather than total restriction
The bank story matters because agents are gaining access to real systems. Feminella says nobody can know in advance exactly what a non-deterministic agent will do, but humans are non-deterministic too. Firms do not prevent human workers from acting or require every person to double-check every other person forever. They give people boundaries, visibility, and controls. He rejects the idea that safety and innovation must be traded against each other. A sandbox with no meaningful system access is safe but not useful. Unrestricted agents have weak isolation, and policy text inside prompts is difficult to enforce at runtime.
An agent registry makes scattered systems visible and governable
Agents appear in many forms, including chatbots, MCP servers, notebook experiments, and web tools. Without a registry, it is difficult to know who owns an agent, which tools it can use, what it should do, or whether an existing agent can be reused. Feminella warns that a single approval gate for every agent creates a queue where simple experiments wait behind high-risk systems. Teams may then abandon the process and move to less visible paths. The registry records an agent's tools, endpoints, product, and owner. Approved building frameworks can register this information automatically, while manual registration is also possible.
Review should follow the risk and purpose declared by each agent
A ticket summarizer and a production deployment agent do not create the same risk, so they should not receive the same review process. Feminella says a detailed manual process can leave low-risk agents waiting for weeks or months while reviewers spend their attention on routine requests. Most enterprise agents have a specific, well-understood purpose. That purpose can be described through declared tools, endpoints, and controls. A research summarizer might need a web tool, while a release deployer has no reason to use it. The agent's declared needs can therefore guide both review and access without making every team request permission for every tool.
Runtime permissions should be translated from the registry into enforceable controls
Feminella distinguishes between an agent that reads a public web page and one that uses a role account to call an internal API. Giving every agent the same short allow list would block useful work, while allowing broad access would create unnecessary risk. Instead, agents declare their purpose and the tools they need. When an agent runs in a managed environment, the system reads its registry policies and translates them into enforceable statements about endpoints, hosts, and other resources. Teams do not need to file a ticket for every access change. If the agent is registered and stays within its risk score, the access policy governs it automatically.
Shared evaluations give agents a common testing layer
Teams often build different evaluation systems, and some build none. Feminella compares evaluations to unit tests because a model change is a runtime dependency that can alter an agent's outputs and quality. He rejects a central process that requires a full evaluation suite and committee approval for every prompt tweak or one-line change. That would slow normal iteration. Instead, the firm should provide a shared evaluation approach and common infrastructure. Teams then add the tests and measurements that describe their own agent. The common layer covers features that agents have in common without forcing every team to recreate the evaluation stack.
Classifiers can reduce approval fatigue by routing only meaningful decisions to people
Human approval for every agent action fails at production volume. Reviewers can receive requests faster than they can assess them, then start approving reflexively. Feminella describes an example in which an inbox agent lost the instruction to wait for approval after compacting its context and deleted hundreds of emails. His alternative uses independent, conservative classifiers for routine calls. A request that is approved almost every time may belong in a rule rather than a human approval queue. Ambiguous or potentially dangerous actions should reach a person, while clearly dangerous actions should be denied. Explicit rules can take a fast deterministic path, with classifiers handling the remaining uncertainty.
The five controls form one operating model for agent autonomy
Feminella combines the controls into a sequence. The registry makes every agent visible from creation. Risk scoring sends human review to the appropriate place. Guardrails limit each agent to the systems and tools it needs. Shared evaluations give teams a common quality bar. Runtime classification makes sure decisions happen inside the operating system for agents rather than through an informal process elsewhere. He compares these controls with car brakes. Brakes do not exist only to stop a car. They allow a driver to travel faster while retaining a way to stop safely. His point is that controls can support useful autonomy when they are composable and applied at the right level.
"If you're saying yes 99.99999% of the time, then that's probably just a rule that the agent should follow, not a request you need to ask the human for permission for."17:29
Who should watch
You are building enterprise agents and need a way to give them useful system access without approving every action manually.
Your organization has many agent prototypes but cannot consistently answer who owns them, what they can access, or how risky they are.
Reviewers are overwhelmed by low-risk requests and need a runtime process that sends people only ambiguous or dangerous actions.