When a security incident begins, responders must work out what is happening, how far it has spread and what to do, usually with incomplete information and under time pressure. Much of the work consists of gathering evidence from many systems, correlating it and choosing actions from known playbooks. These are tasks an agent can assist with. A planning loop lets it examine an alert, decide what to look at next, run the corresponding queries, update its picture and propose a response, repeating until the situation is understood. The promise is faster, more consistent response. The caution is that the same agent, given unchecked authority, can disrupt the business as thoroughly as the attacker, so design focuses on control as much as capability.
How does the planning loop operate in an incident?
The loop begins with an alert and its context. The agent forms a hypothesis, for example that a compromised account is being used for data collection, and chooses evidence that would confirm or contradict it, such as authentication logs, endpoint events and network connections. It runs those queries through tools, reads the results and adjusts. If the hypothesis is supported, it assembles a timeline and a list of affected assets, and proposes containment steps drawn from the organisation's playbooks. If the evidence points elsewhere, it changes course. Good implementations keep an explicit record of each step and the evidence behind each conclusion, so that a human can follow the reasoning and so that the investigation can be reviewed afterwards.
What role does human control play?
Actions differ in their consequences, and oversight should follow. Searching logs and summarising findings carry little risk and can run automatically. Containment steps, such as disabling an account, isolating a server or blocking an address range, can interrupt legitimate work, so they typically require approval from an analyst who sees the proposed action, the evidence and the likely impact. For some well-understood situations, an organisation may allow automatic containment of low-value assets, with limits on how many actions the agent can take in a given period. The agent should be easy to pause and override. Approval that is requested too often becomes a rubber stamp, so thresholds are tuned to keep human attention for decisions that matter.
Evaluating tool use
An agent is only as safe as the tools it can call, so tools should be narrow and well described, with inputs validated and every call logged. Teams evaluate the whole system on recorded or simulated incidents, measuring whether the agent chose useful queries, reached correct conclusions, proposed appropriate actions and recognised when it should escalate. They also test for failure and abuse, since an attacker may plant misleading content in logs or messages that the agent will read, hoping to redirect it. Treating retrieved data as untrusted, separating instructions from evidence and monitoring the agent's own activity like that of a privileged account are essential safeguards. Regular re-evaluation is needed after changes to models, prompts or tools.
Why it matters
Defenders are short of time and skilled staff, and attackers move quickly, so any credible way to speed up investigation is valuable. Agentic AI offers one, with benefits that depend on how carefully it is bounded. Narrow tools, explicit approval of consequential actions, thorough testing and a full record of what the agent did allow organisations to gain speed without giving up accountability.
In an incident the agent should be fast at looking and slow at acting.