Rocket's mainframe agent must clear a policy check before every action it proposes

A retail operations team knew that a production CICS region had stopped after exhausting temporary storage resources, and did not know what had caused it ^[13]. EVA concluded the event was a localized, application-driven issue, isolated the main source of activity, identified the contributing systems and tied the failure to a workload pattern and a likely application owner ^[13]. EVA returns a name to call and the evidence behind the conclusion ^[3].

Buckellew said an operations team at a large South American financial institution spent about three weeks investigating a production problem, and that EVA found a probable root cause and supporting evidence in less than a day after it received the relevant System Management Facilities records and operational context ^[12]. Three weeks is 21 days, so the gap on the step EVA performed is at least 21-fold ^[16]. The comparison leaves out whatever the humans spent locating and extracting the SMF data in the first place ^[12].

PlanGuard is the part that falls to the security team. It examines the caller, the request, the session, the selected tool, environmental conditions and organizational rules when an action is proposed, and it can permit the action, deny it, or require another approval ^[5]. "This invocation-time approach is important for agentic AI because decisions can no longer rely solely on permissions granted during account provisioning," Buckellew said ^[8]. Those organizational rules have to exist per tool before the agent proposes anything ^[5]. Rocket said PlanGuard leaves RACF, ACF2 and Top Secret in place as the underlying enforcement framework ^[7].

The record is built for the person who has to answer for what the agent did. It captures who initiated the request, what the agent proposed, which policy was applied, whether approval was necessary, the identity used for execution and the resulting action, in a tamper-evident, hash-chained trail ^[9]. "This creates a complete chain of evidence that allows operators, security teams and auditors to understand not only what occurred, but why it occurred, with which identity, and under which policy decision," Buckellew said ^[10].

The agent's permissions divide into two columns: what it may propose, and what it may execute. Both cases Rocket described sit in the propose column ^[12][13], and EVA's shipping capability is investigation, with unsupervised system changes still out of scope ^[3]. The short-lived identity and the hash chain are built for the execute column ^[6][9]. Rocket called EVA 2.0 forthcoming and did not give a ship date or a price ^[15].