An agent runs for two hours. Three weeks later, an auditor asks why it made an unusual trade. Your logs show the time and the tool calls. They do not show which policy applied. They do not show who had the authority to approve an exception.
While banks plan longer agent runs, product leaders must design work around them. The staffing question needs to start with exceptions and decision rights. One person must have the authority to stop or redirect the agent. Each decision must also leave a record that stays after the run.
In June, Derek Waldron, Chief Analytics Officer at JPMorgan, told CNBC that the bank planned agents that work without human input for one or two hours. Earlier agents usually ran for two or three minutes. He expected future agents to remain coherent for days, then weeks, according to PYMNTS.
With longer runs, the person reading the agent output becomes the bottleneck in the process. That person must understand what occurred, identify the exception, and decide on the next step. During a longer run, the agent has more time to route work, assign resources, and change records before a person examines them.
When I design an automated workflow, I focus on the handoff between the automation and the people supervising it. I want to know three things: what causes the handoff, what information arrives with it, and which decision the person can make. An exception queue is useful only if the person can resolve the work in it.
A release check cannot supervise an agent that is in operation. The agent continues to make decisions after deployment. The approval at release remains unchanged.
More reviewers, same exposure
Many teams first add more reviewers. But when a person approves of actions all day, that person can stop seeing the important exception. Approval becomes a routine. The process still shows a human checkpoint, but the checkpoint has less value.
The role description must name the decisions that its owner can make. If a reviewer cannot stop or redirect the agent, more reviewers only give more coverage of the same exposure.
The bank job data shows why this point is important. Draup counted 139,819 AI-related job postings in 2026, an increase of 49%, according to PYMNTS. Mentions of agent orchestration increased by 1,721%, from 108 in 2025 to 1,967 in 2026.
Governance terms increased faster than the postings that contain them. Mentions of responsible AI increased by 657%. Mentions of AI governance increased by 394%. BERI’s analysis of the Draup data shows approximately 16,000 governance references. That is approximately two times the number of references to model training, deployment, and operations. It is also eight times the references to agent orchestration.
These numbers count skill mentions in public job ads. They do not count hires. Some terms describe real tasks. Other terms are standard text that companies add to many ads. The important question is this: will the people in these roles have authority over agents in production?
Ivanti’s 2026 research shows the same gap. Of the IT professionals in the survey, 85% said that they have AI policies or oversight procedures. Only 42% said that accountability for AI decisions is clear. A policy can exist even when the responsibility remains unclear.
My first check is the list of agents that operate in the workflow. One large organization estimated 25 to 30 agents and found more than 100. Many of these agents came with vendor software upgrades. If the staffing plan includes only the agents that your team built, it does not include all of the work.
Put the checkpoint at the risky action.
The risk of an action shows where a person must make a decision. Runtime shows how long the agent works. It does not indicate whether the next action requires approval.
According to a June 2025 report, BNY deployed agents to identify and fix system vulnerabilities. The agents then ask a human manager for approval, creating a checkpoint that separates two steps: the agent prepares the patch, and the manager permits its application.
I look for this separation in a workflow design. The agent can prepare the work. The manager owns the decision to apply the change. The human role has a clear purpose, and the manager does not approve each of the earlier requirements.
For most work, use escalation. The agent continues in its permitted scope. When the inputs conflict or the confidence is low, the agent gives the task to a person. Put pre-approval at irreversible actions and at actions that the rules tell a person to approve. Examine low-risk results after the action. Crucial operations and the first weeks of a new agent can call for real-time supervision.
Permissions also set what the human receives. Cisco is giving its MyAgent personal agents to approximately 90,000 employees. According to The Wall Street Journal, each agent has read-only access to one employee’s data. Agents cannot call other agents directly. A person must approve each external action. These limits set what the agent can do before a supervisor receives an exception.
The scope of the agent can increase when measured results show that more autonomy is safe. If the agent does not operate as expected, decrease the scope. Good results in the past do not make a permission permanent.
When you set these limits, you can calculate the staffing:
Daily oversight demand = exceptions per run × runs per day × minutes to resolve each exception.
Compare this demand with the minutes that the team has for oversight each day. Include mandatory approvals in the workload. If demand exceeds capacity, reduce the agent’s scope or increase the team’s capacity. More people in an approval queue without clear limits do not correct the design.
Keep the rule with the decision.
Supervision must leave evidence that explains consequential decisions after the agent stops. Logs can show the time of an action and the account that did it. Logs usually do not show the rule that controlled the decision. They also do not show who had the authority to approve an exception.
For each consequential decision, keep a record of these items:
- The policy version that applied
- The evidence that the agent examined
- The approval authority
- The person who approved, if a person approved the action
- The procedure that the team used for an exception
- The result
Three weeks after the trade, the current policy cannot explain an earlier approval if the rule changed. The decision record must keep that connection.
I also examine where the policy has effect. Policy as code can control configuration and release. If these are its only control points, the check does not control the later actions of the agent. During a market event, a team can tighten its risk threshold. But the agent does not get the new threshold until the next pipeline run.
Banking regulators gave this agent work to the banks. On April 17, 2026, the Federal Reserve, OCC, and FDIC replaced the 2011 model risk guidance. The new guidance does not include generative and agentic AI. The guidance tells banks to use their own risk management and governance practices for the systems that it does not include.
If there is no record, the liability does not move to the agent. The company stays liable. In Moffatt v. Air Canada (2024), a Canadian tribunal rejected the airline’s argument that its chatbot was responsible for its own statements.
The oversight team has two tasks. It resolves the exceptions in the queue today. It also keeps records that allow the owner to explain a decision weeks later.
Check your plan before you add headcount.
Start with question zero: do you have a current list of the agents in this workflow? Include the agents that came with vendor products.
Then answer four questions for each agent:
- Which actions must have pre-approval? Which conditions cause escalation? Which results can people examine after the action?
- What is the daily workload determined by the exception frequency, the number of runs, and the time to resolve each exception? Can the team do this work?
- Who has the authority to approve, redirect or stop the agent?
- Can you rebuild a consequential decision from its policy version, approval record, exception procedure, and result?
Use gaps as gates. If the list of agents is incomplete, some work is not included in the staffing estimate. If demand exceeds capacity, reduce scope or add staff. If a reviewer does not have authority, that reviewer cannot stop the agent. If there are no records, the owner cannot explain the agent’s actions.
If no person will put a name next to the consequential actions of an agent, that agent is not ready to run for two hours.