Agencies should evaluate agents like they would evaluate a team. How much time are the agents spending improving themselves, collaborating, or duplicating work? Does each agent’s independent approach create negative consequences? Or is the work getting done effectively enough that some duplication does not matter?
Observability for agents is really about explainability. Can I understand, in human terms, what they are doing well and how that benefits the agency? Can I understand what they are doing poorly and the consequences for the mission?
Agencies also need a view of the entire workflow, including humans and agents. I don’t want a PDF explaining what happened last month. I want a real-time dashboard that lets me inspect activity at a low level and build a broader mission picture.
I also want anomaly detection. Agencies will not have enough human eyes to monitor every agent operating in a high-volume cyber or mission environment. The people, processes, and technology supporting the workflow should create a self-correcting function that identifies irregularities in real time.
MeriTalk: What factors should agencies consider when deciding which decisions agents can make independently and which require human review?
Erickson: Agencies should begin with a concept of operations. Before deciding what to automate, they need to understand how the workflow should operate and what role they want people to play.
There is an important distinction between human in the loop and human in the lead. Human in the loop means the agent performs a series of actions, and the human is passive until the completed package of work arrives for review.
Human in the lead is proactive. Imagine a social worker who is trying to achieve the best outcome for a case. Instead of waiting for a package of work or an alert, that person can ask agents to find services, gather more context, or double-check information. AI creates an opportunity to support the people who are actively directing the work.
Some deterministic decisions may not need a human at all. For non-deterministic decisions that require judgment, the person and the AI should work together, with the human in the lead.
Agencies still need to consider risk, the consequences of an incorrect decision, and whether the decision process can be codified and explained. But the larger question is not simply whether we are willing to give the robot the keys to the car, for example. It is why we are driving somewhere in the first place, who selected the destination, and whether that decision supports the mission.
MeriTalk: How can agencies evaluate whether AI agents are improving workflow speed, accuracy, consistency, and mission outcomes?
Erickson: The federal government is in a unique position because it has spent decades acting as a regulator. Agencies routinely ask banks, manufacturers, and critical infrastructure organizations to show how they reached a decision, what they knew at the time, and whether they followed an established process.
Government should take the same approach with its AI agents: Show me how you did it. Retain the context behind the decision so it can be reviewed later.
The questions are whether the agency is keeping the data, whether that data is trapped inside a proprietary black box, and whether authorized reviewers can query it in real time without breaking the bank. Agencies need standards-based mechanisms that allow them to establish the ground truth and measure outcomes against it.
Those measurements will evolve because policies will evolve. Agencies need an ongoing process for adapting their metrics while preserving the evidence they need to audit decisions.
We already do this in cybersecurity and IT systems management. When a ransomware event occurs, we know how to investigate what happened. Agencies should build that same level of preparedness for situations in which an agent makes a bad decision.
MeriTalk: As agencies expand agentic AI across high-volume workflows, how can they keep human attention focused on the decisions and exceptions that matter most?
Erickson: I don’t think the industry has one finalized answer. One approach Elastic is exploring is the use of higher-level agents that look after other agents and work side by side with cyber and IT operations professionals.
The human-computer interface matters. I don’t want to pair a cyber defender with a deeply technical coding agent that operates at an entirely different level. I want to pair that analyst with an agent that understands the analyst’s role, communicates at the same level, and knows how to direct more technical agents behind the scenes.
Agencies should also ask how to take operational burdens away from their best people. Too much human attention is still focused on racking and stacking servers, maintaining infrastructure, or determining how many cloud servers to spin up. Turnkey, serverless, autoscaling, and FedRAMP-authorized services can help agencies shift that attention toward emerging threats, mission decisions, and AI adoption.
We are already seeing organizations move quickly in cybersecurity because adversaries are using machine-speed threats and agents. Defenders need agents as well. The technology is available. The question is whether agencies can build the processes needed to absorb it and keep their people focused on the mission.
