I had the pleasure of joining several leaders at the Agentic SOC Forum: From SOC Cacophony to Agentic Neural SOC Symphony.
The experience of the leaders I joined on stage spanned national cyber defense, enterprise security, healthcare, cloud data, public-sector technology, academia, manufacturing, and business transformation. What came out of the conversation at Atlanta Tech Park that day was a clear picture of the future of the security operations center (SOC). Ideas were challenged. Tested. And overall, the bar was raised.
The SOC has spent years adding technology in pursuit of greater speed, visibility, and control. Security teams automated repetitive tasks, connecting tools through orchestration platforms. They deployed artificial intelligence to correlate alerts, summarize incidents, and recommend actions.
Each advancement created value. None of them, however, fully changed the operating model.
Agentic AI does.
When security leaders hear the term agentic SOC, they shouldn’t think of another automation layer or a more sophisticated chatbot. They should think about a fundamental shift in how security work is initiated, coordinated, executed, governed, and improved.
Traditional automation follows instructions. Agentic AI pursues objectives.
That distinction may sound subtle, but it changes everything.
From Following Playbooks to Pursuing Objectives
Traditional security automation is deterministic. A condition occurs, a rule is triggered, and a predefined action follows. If an alert meets certain criteria, the system enriches it, creates a ticket, sends a notification, or initiates another prescribed task.
This approach is highly valuable because it’s predictable, repeatable, and easy to audit. Its limitation is that it can only address scenarios anticipated by the people who designed the workflow. Security orchestration expanded that model. SOAR platforms connect multiple products and coordinate more complex playbooks across identity, endpoint, network, cloud, ticketing, and communications systems. Orchestration made it possible to automate an entire chain of activity rather than one isolated task.
But the decision tree still had to be designed in advance.
AI-assisted security tools advanced the model again. They can help analysts correlate signals, summarize investigations, prioritize alerts, generate queries, identify patterns, and recommend next steps. These capabilities reduce cognitive load and help people make decisions faster.
Yet the human analyst generally remains responsible for directing the investigation and determining what happens next.
An agentic SOC crosses a different threshold. An AI agent can be given an objective, determine a course of action, use approved tools, evaluate the evidence it collects, revise its hypothesis, and continue working without requiring human direction at every step. Recent enterprise governance research similarly characterizes autonomous agents as systems that perceive, decide, and act, rather than tools that only provide decision support.
The distinction can be summarized simply: Automation executes tasks. Orchestration connects workflows. AI assistance improves human decisions. Agentic AI performs bounded decision execution in pursuit of an objective. That is not merely a technological improvement, but a reassignment of operational authority.
What an Agentic Investigation Looks Like
Consider an unusual identity alert. A traditional automated workflow might enrich the alert, check the IP address against threat intelligence, open a ticket, and notify an analyst. An AI-assisted tool might summarize the user’s recent activity, highlight related endpoint signals, and recommend that the analyst investigate possible credential compromise. An agentic workflow can go further. Within its approved authority, an agent could:
- Gather identity, endpoint, network, cloud, and threat intelligence evidence.
- Form an initial hypothesis about the event.
- Test that hypothesis against related activity.
- Look for indications of persistence or lateral movement.
- Evaluate the reliability and completeness of the evidence.
- Document its reasoning and decision path.
- Recommend or initiate a containment action based on predefined risk and confidence thresholds.
- Escalate the incident when its evidence is inconclusive or the proposed action exceeds its authority.
This is where the promise becomes real, but it’s also where the risk changes.
A system that summarizes an incident creates one category of risk. A system that can disable an account, isolate an endpoint, block network traffic, or modify a production environment creates another. Security leaders therefore must stop asking only, “How accurate is the model?”
They must also ask:
- What is this agent permitted to do?
- Which systems and data can it access?
- Who approved that authority?
- What actions require human approval?
- What evidence must it preserve?
- How is the agent monitored?
- How can its actions be reversed?
- Who is accountable when the agent is wrong?
These are questions of operating authority, not simply model performance.
The Biggest Challenge Is Not Capability. It Is Trust.
We have largely moved beyond asking whether AI agents can perform meaningful SOC tasks. Many organizations can demonstrate an agent that triages alerts, collects evidence, runs queries, or recommends containment actions. The harder question is whether security leaders trust an agent enough to act. Piloting an agent in a controlled environment is one thing. Granting its authority inside a production environment is something entirely different.
Trust cannot be established through a compelling demonstration. It must be built through explainability, auditability, repeatability, controlled permissions, and consistent outcomes. The internal panel framework captured this point directly: the future SOC will be defined not only by the capability of its AI, but by the confidence security teams place in AI-driven decisions.
That confidence requires evidence.
Leaders need to know what data the agent used, what assumptions it made, what tools it called, what alternatives it considered, why it selected an action, and whether the outcome matched expectations. If the agent changes course during an investigation, that change must be visible. If it escalates or takes action, the reason must be re-constructable. A black box cannot become a trusted security operator.
Human-in-the-Loop Is the Bridge to Adoption
Some organizations treat human approval as an obstacle that eventually needs to be engineered out of the process. I believe that’s the wrong way to view it. Human oversight is not a weakness, but a mechanism through which responsible autonomy is earned. The strongest starting point is usually recommendation, not unilateral control. The agent investigates, develops a conclusion, recommends an action, and presents its evidence to an analyst. The analyst validates, rejects, or modifies the recommendation.
Every one of those interactions creates a learning loop. The organization learns where the agent performs reliably, where the ambiguity remains, which controls are effective, and which scenarios should stay under human authority. Over time, repetitive and low-risk actions can move from recommendation to execution with approval. Once reliability is established, selected actions may become autonomous within clearly defined boundaries. The path to autonomy begins with augmentation.
This is consistent with the broader human-centered approach I advocate for regarding AI adoption. The objective is not to remove people from meaningful work. It’s to remove repetitive processing that prevents skilled people from doing meaningful work.
In the SOC, that means giving analysts more capacity for threat hunting, incident strategy, detection engineering, governance, architecture, and complex judgment. The agent handles the operational burden it has demonstrably earned the right to handle and humans retain authority over uncertainty, material risk, and exceptions.
Governance Must Evolve from Model Governance to Agent Governance
Most enterprise AI governance programs were designed around models that generate predictions, classifications, or content. Those programs often focus on accuracy, privacy, bias, data quality, transparency, and acceptable use.
Those controls remain important, but they are not sufficient for systems that can act. Agent governance must address the complete decision and execution chain. It must define identity, access, permissions, scope, monitoring, escalation, intervention, evidence retention, reversibility, and accountability.
In practice, leaders should establish an authority envelope for every agent. That envelope should specify:
- The objectives the agent may pursue.
- The systems, tools, and data it may access.
- The actions it may recommend.
- The actions it may execute with approval.
- The actions it may execute autonomously.
- The conditions that require escalation.
- The confidence and risk thresholds governing each action.
- The evidence the agent must retain.
- The controls available to pause, override, or terminate activity.
- The human owner accountable for the agent’s operation.
This is particularly important because agents function less like passive software and more like digital operators with varying levels of privilege. External governance research has warned that autonomous systems can introduce both new access points and new internal risks when they operate across connected enterprise systems.
The principle should be clear:
Autonomy does not eliminate accountability. It increases the need to define it.
An organization cannot delegate responsibility to an algorithm. Humans remain accountable for deciding where agents operate, what authority they receive, how they are supervised, and how failures are addressed.
The Talent Gap Is a Catalyst, Not a Justification for Replacement
The momentum behind agentic security is not simply the result of enthusiasm for new technology. Security teams face a capacity problem.
The volume, speed, and complexity of security activity challenge the ability of human teams to process every signal with the depth it deserves. Internal forum material described the opportunity as moving from fragmented alerts and disconnected tools toward coordinated, accountable security operations guided by human command. Agentic AI offers a way to increase investigative and response capacity without increasing staffing at the same rate.
But leaders must be careful about how they frame that opportunity.
If agentic AI is introduced primarily as a headcount reduction strategy, organizations may create resistance, weaken adoption, and lose the expertise required to supervise the agents effectively. The better approach is to treat AI as a force multiplier for cyber talent. Analysts should spend less time copying information between tools, enriching routine alerts, repeating standard queries, and documenting predictable steps. They should spend more time improving detections, investigating novel threats, refining agent behavior, validating complex conclusions, and strengthening organizational resilience.
The future role of the analyst is not simply alert handler. It increasingly includes:
- Agent supervisor
- Decision validator
- Exception manager
- Threat strategist
- Detection engineer
- Governance participant
- Automation and agent designer
This is a redesign of the work, not the removal of the worker.
Adopt Through a Crawl, Walk, Run Model
Organizations should resist the temptation to pursue full autonomy immediately. The fastest route to trusted deployment is disciplined progression.
Crawl: AI-Assisted Investigation
Begin with bounded use cases such as alert summarization, evidence gathering, enrichment, investigation recommendations, and documentation support. At this stage, the agent recommends. The human decides. The goal is to evaluate the quality of reasoning, evidence, consistency, and escalation behavior without introducing material operational risk.
Walk: Semi-Autonomous Execution
Once reliability is demonstrated, allow the agent to initiate selected actions with analyst approval. The agent may prepare an endpoint isolation request, recommend an account suspension, draft a block rule, or assemble a remediation plan. A human reviews the evidence and authorizes execution. The goal is to prove that approval thresholds, audit trails, escalation paths, and recovery procedures work under real operating conditions.
Run: Bounded Autonomy
Only after the organization has established confidence should an agent independently execute low-risk, reversible, and well-understood actions. Autonomy should expand based on evidence, not ambition. Success in one use case does not automatically justify authority in another. Each capability should earn its own level of trust. This progression was central to the attached panel framework: trust is earned incrementally, not granted immediately.
What Security Leaders Should Do Now
Leader’s shouldn’t begin looking for an “agent” that can do it all. Instead, It’s identify one recurring SOC problem that is:
- Operationally meaningful
- Narrow enough to govern
- Supported by reliable data
- Measurable
- Reversible if something goes wrong
- Currently consuming disproportionate analyst time
Then define the objective, authority envelope, evidence requirements, human checkpoints, success measures, and escalation conditions before deployment. Measure more than speed. Ask whether the agent improves investigation quality, reduces unresolved work, preserves evidence, makes consistent decisions, improves analyst capacity, and produces outcomes the organization can defend.
Most importantly, treat deployment as an operating model initiative rather than a technology installation.
The Leadership Imperative
The agentic SOC will not be defined by how many agents an organization deploys. It will be defined by how responsibly those agents operate alongside people. The winners will not be the organizations that move fastest toward autonomy. They will be the organizations that build the strongest foundation for it.
They will understand that:
- Capability without governance increases exposure.
- Autonomy without auditability weakens accountability.
- Speed without judgment creates new risk.
- Human oversight is part of the architecture, not evidence that the architecture failed.
- Trust must be demonstrated before authority is expanded.
The future SOC will not be human versus AI. It will be human expertise amplified by trusted AI agents. Analysts will provide judgment, creativity, context, and accountability. Agents will provide speed, persistence, scale, and the ability to coordinate complex investigative work across systems. That combination is the real promise of agentic security operations.
Not fewer people. Not more tools.
A more intelligent, coordinated, and accountable security operating model where people and agents each do what they do best.