What the CrowdStrike Agentic SOC Means for Security Operations

The CrowdStrike agentic SOC is a set of AI agents in the Falcon platform that split up the work of investigating a security alert and share what they find. Each customer sets limits on what the agents can do. CrowdStrike announced the latest version at its Fal.Con 2026 conference on September 2. The coordinated investigations at the center of the announcement are pre-beta.

For a team that runs Falcon, the practical question is which parts of CrowdStrike's agentic SOC it can use this quarter. The answer depends on what CrowdStrike announced, which parts are available today, what the agents can see outside Falcon, and who controls what they do.

{{ebook-cta}}

What CrowdStrike announced at Fal.Con 2026

CrowdStrike's press release describes the next version of its agentic SOC this way:

• Coordinated investigations. Charlotte AI sends domain agents to work an investigation in parallel. CrowdStrike says investigations "cover every domain simultaneously, including threats" against a company's own AI systems, such as model abuse, prompt injection, and data theft through AI assistants.
• Shared context layer. It builds on CrowdStrike's Enterprise Graph. In the release's words, "Every agent shares one memory of the environment. What one learns, they all know."
• Certified data pipelines. Built on Falcon Onum, they carry third-party data into Falcon Next-Gen SIEM and are certified by Falcon Complete, CrowdStrike's managed detection and response service. CrowdStrike says they can cut data storage costs by up to 50 percent. It also says "detection runs inside the pipeline, so threats are caught in-stream before data ever lands."
• Charlotte Agentic SOAR. It gives teams one workspace to build and govern automation. The release says "customers set the autonomy level for each workflow, from human-in-the-loop approval to fully autonomous execution."
• Bidirectional MCP. It "connects any third-party agent into Falcon and any CrowdStrike agent" out to external tools.

The release also notes that it "includes discussion of unreleased services or features," without saying which ones. CrowdStrike's launch blog labels the status of each.

Which parts of the CrowdStrike agentic SOC are available today

As of September 18, 2026, three of the eight capabilities listed for the CrowdStrike agentic SOC are pre-beta, including the coordinated investigations and certified data pipelines. Two are in Public Preview, two are generally available, and one is labeled Public. The launch blog lists them this way:

Source: CrowdStrike's launch blog, updated September 18, 2026.

The blog gives no release dates for the pre-beta items. The press release adds that the unreleased features it describes are still in development and subject to change. For now, the generally available pieces are model choice and MCP connections.

What the CrowdStrike agentic SOC can see outside Falcon

According to the blog, investigations in the CrowdStrike agentic SOC will run in Falcon Next-Gen SIEM on existing Falcon telemetry, with no new sensors. So the agents start with Falcon's endpoint, identity, and cloud data.

How far that gets an investigation depends on where the alert started. A Falcon EDR alert begins with Falcon's own data. An Okta alert needs the user's sign-in history from Okta, and a phishing report needs the original message and the email gateway's verdict. Both depend on data from other vendors' tools. While the coordinated investigations and pipelines are pre-beta, a team can't yet measure how those investigations handle either case on its own alerts.

That outside data can reach the agents two ways: through Next-Gen SIEM, or through MCP connections out to those tools. For Next-Gen SIEM, the blog describes a choice for each source: "Teams decide what is ingested and what is federated, and critical first-party data has no ingestion cost." Ingested data is copied into Next-Gen SIEM, while federated data stays in its own store and is queried there. The certified pipelines add CrowdStrike-maintained routes for specific sources, starting with Zscaler and Palo Alto Networks. A week after Fal.Con, Zscaler launched an agentic SOC of its own, built on Zscaler's telemetry plus third-party data. So a team that runs both now has two vendors offering to investigate the same Zscaler data.

For ingested data, CrowdStrike offers Insight XDR customers 10 GB a day of free third-party data ingestion into Next-Gen SIEM. The offer page gives no further terms, and neither the press release nor the blog gives a price for third-party ingestion. Microsoft ISOC follows the same pattern: it includes Defender data and charges $2.40 per gigabyte for data from other vendors. In both cases, data from other vendors has to be priced separately.

MCP also works in reverse, bringing outside agents to Falcon's data. The generally available "connect to any tool, any agent" capability lets third-party agents reach Falcon tools through a CrowdStrike-managed MCP server. So a team's agents from other vendors can work with Falcon data now, while CrowdStrike's own multi-agent investigations are still pre-beta.

How autonomy works in Charlotte Agentic SOAR

The autonomy setting in Charlotte Agentic SOAR covers who approves an action, from a human approving each one to fully autonomous execution. Who decides the steps is a separate question.

Anthropic, whose models CrowdStrike customers can bring under their own licenses, separates two kinds of automation in its guide to building effective agents. It describes workflows as "systems where LLMs and tools are orchestrated through predefined code paths." Agents are systems where LLMs "dynamically direct their own processes and tool usage." So a fully autonomous workflow can still follow a path someone wrote in advance. An agent chooses its next step at runtime, based on what it has found so far.

Charlotte Agentic SOAR is built to hold both kinds. The blog describes the workspace as one place to build and govern rule-based and agentic automation, with control over triggers, agents, actions, and which steps need approval. When a team reviews a workflow, it should check which kind it is. A predefined path has no step for a case its author did not anticipate.

Per-workflow settings let a team require approval on every action at first, then remove it for actions that have proved reliable. Picking which actions to hand over first is the same decision about what to automate that a team faces with any vendor.

How the CrowdStrike agentic SOC answers five agentic SOC questions

Five questions apply the line between workflows and agents to a SOC product, and they work for any agentic SOC platform on a shortlist. Here is how CrowdStrike's public materials and Prophet AI answer them:

Sources: CrowdStrike's press release, launch blog, and Charlotte AI triage announcement; Prophet Security's AI SOC Analyst and integrations pages.

What to check before the CrowdStrike agentic SOC becomes the SOC plan

Before the CrowdStrike agentic SOC goes into next year's plan, the security team should answer ten questions:

• Status. Which capabilities in the plan are GA, which are in Public Preview, and which are pre-beta? What dates will CrowdStrike put in writing?
• Alert sources. What share of the team's alerts comes from tools other than Falcon, and can the agents investigate those alerts today, at every severity, without waiting for a certified pipeline?
• Connector depth. For each non-Falcon tool, can the agents query it for more evidence during an investigation, or only see what was copied into Next-Gen SIEM?
• Cost. What will it cost to put the team's non-Falcon data in front of the agents, beyond the 10 GB a day of free ingestion that comes with Insight XDR?
• Triage or investigation. Do the agents only sort alerts into true and false positives, or run a full investigation on each one?
• Accuracy. On a sample of the team's own past alerts, how often do the agents reach the same verdict as the team's analysts, and how many real threats do they miss? CrowdStrike's published accuracy figure for Charlotte's triage, over 98%, measures agreement with CrowdStrike's own MDR analysts.
• Adaptability. Can analysts change how the agents investigate, such as adding the team's own playbooks and business context, without building and maintaining their own agents? Can they see each change and roll it back?
• Evidence. For a sample of agent verdicts, can analysts see every query the agents ran and run it again themselves? The explainability test applies to every vendor.
• Data use. Does any of the team's data, or its analysts' verdicts, train CrowdStrike's models? Under bring your own model, what does the model provider keep?
• Lock-in. If the team replaces its EDR or SIEM later, what happens to its investigations and automations?

A proof of value on the team's own alerts can test the GA pieces now. For the pre-beta pieces, the tests used to judge any agentic SOC platform apply once CrowdStrike can run them on the team's data.

What the CrowdStrike agentic SOC means for security operations

As announced, the CrowdStrike agentic SOC puts agents to work together on Falcon data, with data from other vendors reaching them through Next-Gen SIEM and MCP. For a team whose alerts come mostly from Falcon, the agents will work on data the team already has. For a team with a mixed stack, the plan depends on the coordinated investigations and certified pipelines, and both are pre-beta with no public dates.

Prophet AI investigates every alert across the tools a team already runs, CrowdStrike and non-CrowdStrike alike. See how it works on your own alerts.