Read the order of operations before the feature list. Guardian has to see an agent before Agentic IdP can register it, and an agent without a registered identity is not eligible for authorization at all [6][7]. In practice that means one of two outcomes on Monday: a workflow halts and someone has to track down who owns it, or an exception ticket reinstates the standing credential that token brokering was supposed to eliminate [8].
Most agents in production today run on a service account, an API key or a workload identity set up by whoever was closest to the problem at the time, carrying whatever permissions that person happened to have, even though teams often describe them as running under managed identities [5]. CrowdStrike's own framing is that none of those were designed for software that acts on a person's behalf and hands work to sub-agents [5]. Accept that and the bill lands on engineering rather than security: an agent written to read a key out of an environment variable has to learn to ask for a scoped, short-lived token instead [8], and every action it takes now carries the name of a human or workload that answers for it [9].
The SOC half is a throughput claim. Charlotte AI now runs agents against endpoint, identity, SaaS, cloud and network at the same time on a single investigation, rather than one agent per alert worked in sequence, which CrowdStrike argues returns a fragment from one domain instead of a verdict [11][12]. The part that makes parallel work coherent is the shared context layer on Enterprise Graph, described as persistent memory across agents, investigations and tenants, so what one agent establishes the others can use without a handoff [13]. CrowdStrike puts platform volume at close to 4 trillion events a day across those five domains [14], which works out to roughly 46 million events a second (4,000,000,000,000 divided by 86,400) [16]. Michael Sentonas calls single-agent SOC tooling "table stakes" and cross-domain agents "the new standard" [17]; the hours-into-minutes figure attached to the rebuild is the company's own [11]. More useful for the person evaluating it: investigations now cover prompt injection, model abuse and exfiltration through AI assistants, and the agents return reasoning with the verdict rather than a bare conclusion [15].
The forcing function for next week is a two-column list of every agent you run in production. Column one, can you name the human or workload it acts for. Column two, does anything break if you revoke its standing credential tomorrow. If there's a named owner and nothing breaks, you can register the agent and move on. If there's a named owner and something breaks, that's scoped engineering work with a cost you can estimate. If there's no named owner and something breaks, that agent belongs to the group this whole model was built to handle, a group that typically goes uncounted. If that list comes out shorter than your actual agent count, the discovery product matters more to you right now than the identity product does.
