AI Agents Quietly Replace Human Decision Makers
Somewhere in a satellite operations center, a network engineer used to spend her mornings reviewing anomaly reports. She would read them, weigh them, decide which ones warranted intervention. That was the job. Today, the system still generates the reports, but nobody reads them before the fixes are applied. The engineer’s judgment — the thing she was hired for — has been moved to a review queue that opens after the fact. She is not unemployed. She is superfluous in the way that matters most: the decisions happen without her, and her role is to nod.
The Difference Between Waiting and Acting
A chatbot sits. It waits for a question, produces an answer, and returns to its dormant state. An agent does not wait. It perceives its environment, forms goals, and takes action without a prompt. That distinction sounds technical. It is not. It is the difference between a tool you pick up and a colleague who picks up the work. When multiple agents interact — each perceiving, each acting, each adapting — the result is a system that can coordinate without a coordinator. The human who once sat at the center of that coordination becomes a spectator to a process she no longer steers.
Gartner defines an agent broadly: not necessarily an AI model, but any software program or computational entity capable of perceiving its environment and acting on it. 1 Combine several of these independent but interactive entities, and you have a multiagent system. The definition matters because it reveals how long this has been building. These systems are not new. What is new is what they can do without us.
The Clean Room That Runs Itself
NASA’s Jet Propulsion
Laboratory needed clean rooms — environments where nothing contaminates flight hardware bound for other planets. 1 The previous sensors cost tens of thousands of dollars each and offered extreme sensitivity. The replacement sensors cost hundreds. They were less sensitive, which should have been a downgrade. Instead, engineers networked them into a cooperative system where each sensor agent shared data and knowledge with its peers. Collectively, they decided whether particulate levels were too high. The individual measurement became less important than the collective judgment.
That system still operates today. But the flashpoint came when these agentic frameworks absorbed deep learning and deep reinforcement learning. Instead of following rules and thresholds, the systems now train on data and improve over time on their own. Chris Mattmann, who led the JPL work and now serves as chief data and AI officer at UCLA, puts the risk plainly: with tools like ChatGPT or Midjourney, there is constant human interaction. 2 With agentic AI, the potential for autonomous decision-making removes the human from the loop entirely. The clean room sensors never needed a person to interpret their readings. The newer systems do not need a person to act on theirs either.
Thirty Years of Watching the Human Step Back
Anand Rao, an AI professor at Carnegie Mellon University, worked on multiagent systems in the 1980s. 3 He built them for the Space Shuttle, which was too complex for any single system to manage. The agents of that era were rudimentary by today’s standards, but the architecture was the same: distribute perception and action across multiple entities, let them coordinate. Over four decades, those agents became more independent. They moved from following explicit instructions to acting on goals and objectives. Each step reduced the need for a human to translate intention into action.
The trajectory is consistent. What began as a way to handle complexity — too many variables for one mind to track — became a way to handle decisions. The human who once sat between the data and the response has been gradually pushed to the periphery. Not removed. Just no longer necessary for the system to function.
The Network That Fixed Itself Before Anyone Noticed
Hughes Network Systems, a satellite communications provider, has used agentic AI for years to address service degradation. 4 Dan Rasmussen, the company’s SVP and GM for North America Enterprise, describes a system that continuously feeds network and customer equipment stats into algorithms. 4 The algorithms adapt to changing conditions and identify anomalies. At first, the tools made recommendations. Humans reviewed them and decided what to implement. After a few months of observation, Hughes allowed the process to run automatically and report on changes after they were made. The human review moved from before to after. The decision moved from person to process.
Alarms still exist. If the number of anomalies detected or remediations implemented exceeds statistical norms, the automation stops. But the baseline assumption has shifted. The system is trusted to act. The human is trusted to audit. That inversion — action first, review later — is where a role becomes superfluous. The person is no longer the one who decides. She is the one who confirms that the decision was acceptable.
Hughes has since expanded its use of agentic frameworks, building software to automate application deployment, create CI/CD pipelines, and generate code. The company has completed a proof of concept for analyzing logs and metrics, connecting them with source code to gain insights into code fixes, vulnerabilities, and performance issues. The goal is a product. The implication is that the engineers who once performed those analyses will shift to overseeing the systems that perform them. Their expertise becomes a check on the machine rather than the source of the work.
The Survey That Measures the Shift
A
Capgemini survey of 1,100 executives at large enterprises found that 10% of organizations already use AI agents. 5 More than half plan to use them within the next year, and 82% plan to integrate them within three years. Seventy-one percent said AI agents will increase automation in their workflows. Sixty-four percent said they will improve customer service and satisfaction. Fifty-seven percent said the potential productivity improvements outweighed the risks.
The largest use case is software development — generating, evaluating, and rewriting code. Seventy-five percent of companies planning to use AI agents intend to use them this way. Babak Hodjat, CTO of AI at Cognizant, explains why: most AI researchers are software engineers, and software engineering is easier to verify. 6 You can build semi-supervised systems that check each other’s work. The first human role to be made superfluous is the one that writes and reviews code, because the work is structured enough for machines to validate each other’s output.
That is the pattern. The roles that go first are the ones where the work can be checked without a person. Verification becomes automated. Judgment becomes a statistical threshold. The human who once provided the final word becomes a fallback for when the threshold is exceeded.
The Security Chief Who Cannot Sleep
Dan Meacham, CISO at Legendary Entertainment, uses agentic AI to hunt for movie leaks and piracy across social media and the dark web. 7 The systems leverage open source intelligence. They operate across various verticals in the company’s security programs. Meacham declined to name the specific frameworks, but described an enterprise OpenAI-like solution that enables business process automation.
He also described the risks. Agentic AI is a double-edged sword, he says. “We do lose sleep on this.” [7] The concerns include creative content leaking through AI applications and AI generating infringing content. Then there is the risk of malicious code injections — code hidden inside documents that an AI agent reads and then executes. Meacham notes that this attack vector is not new; it resembles classic SQL injection or database stored procedure attacks. But the context is new. The agent is acting autonomously. The human is not in the loop to catch the anomaly.
Emerging mitigation techniques use data loss prevention patterns to limit or exclude data types from being learned. Some solutions claim to inspect instruction sets as modules learn and grow, preventing injections, hallucinations, and malicious code. By early 2025, Meacham expected real contenders in AI protection — more than just existing DLP and code review enhancements layered on top of language models. The promise is a safeguard. The reality is that the safeguard is another automated system, checking the work of the first automated system. The human is further removed.
The Guardrails That Replace the Guard
Sinclair Schuller, a partner at EY, describes strategies to secure multi-agent AI beyond the guardrails already set up for underlying generative models. 8. An agent can have a particular personality, he says. It can be long-lived or short-lived. It can be allowed to collaborate with other agents or restricted from doing so. Multiagent systems can use consensus, asking peer agents to evaluate each other’s work. They can use adversarial setups, where one agent tries to break another’s output.
Each of these strategies replaces a human function with a machine function. Peer review becomes agent review. Red-teaming becomes adversarial agent testing. The security professional who once performed these evaluations becomes the designer of the system that performs them. That is a promotion for some. For others, it is the moment their hands-on expertise becomes a specification rather than a practice.
The guardrails are necessary. They are also a confession. The more autonomous the system, the more elaborate the checks must be. And the more elaborate the checks, the more the human role shifts from doing the work to designing the constraints on the work. The person who once decided is now the person who decides what the decider is allowed to decide.
The Boundary Between Correlation and Causation
There is a temptation to read these developments as a story of replacement. That is too simple. The more precise observation is that agentic systems are absorbing the judgment that used to reside in a person. The judgment was never pure. It was always shaped by training, experience, and context. The machine learns from data that encodes those same influences. The difference is that the machine does not need to be present. It does not need to be consulted. It acts, and the human learns about the action afterward.
This is where the boundary between correlation and causation becomes blurry. The system identifies an anomaly and applies a fix. The fix works. The correlation between the system’s action and the outcome is recorded. Over time, the system learns which actions produce which outcomes. The human who once made that judgment based on experience is now the source of the training data, not the source of the decision. Her expertise is encoded, then superseded.
The Detail That Shows How Far Practice Is From the Promise
In the clean room at JPL, sensors that cost hundreds of dollars replaced sensors that cost tens of thousands. They were less sensitive individually. They compensated by working together. The system still runs. The humans who designed it have moved on to other problems. The agents continue to perceive, decide, and act.
That is the detail. Not the survey numbers, not the enterprise deployments, not the security concerns. The detail is that the system runs without anyone watching. It was built to solve a problem. It solved the problem. The people who built it are no longer needed for it to continue. That is not a failure. It is the quietest kind of success — the kind where the human role becomes superfluous, and nobody notices because the work still gets done.
Sources
1. NASA’s Jet Propulsion Laboratory
2. UCLA
5. Capgemini
6. Cognizant