Rocket Software brings governed AI agents to mainframe operations
SiliconANGLE Paul Gillin
Rocket Software is adding guardrails to its AI agents for mainframes. The pitch: let AI help with fixes, but not wander into production like a bull in a server room.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Rocket Software is pushing its Enterprise Virtual Assistant deeper into mainframe work, but with a hard stop before full autonomy. The next version, EVA 2.0, will add a security layer called PlanGuard that checks an agent’s proposed action before anything runs. That matters because Rocket is aiming at a tricky middle ground: using generative AI to investigate and eventually automate operational work without handing over unrestricted access to core systems.
Right now, EVA is built for investigation. A person asks a question in natural language, and the assistant can pull together tools and data sources, collect context, compare evidence across systems, and return findings with recommendations. Rocket says the point is to replace the scramble across consoles, dashboards, logs and reports with one explanation that ties the pieces together.
PlanGuard is the mechanism that keeps that from turning into a free-for-all. It looks at the caller, the request, the session, the chosen tool, environmental conditions and company rules before allowing an action. The result can be approval, denial or another human sign-off. If it does approve the task, the system creates a temporary execution identity for that specific job and removes it when the work is done. Rocket says it works alongside RACF, ACF2 and Top Secret, rather than trying to replace them.
The company is also trying to make the machine’s behavior easier to audit. EVA records who made the request, what the agent suggested, which policy was used, whether approval was needed, which identity executed the task and what happened next. It also adds a tamper-evident, hash-chained audit trail so operators, security teams and auditors can trace the chain from question to action.
Rocket is testing EVA with customers in financial services, government, insurance, retail and telecommunications. In one pilot, a South American financial institution spent about three weeks chasing a production issue before EVA narrowed it down in less than a day after being fed relevant System Management Facilities records and operational context. In another, a retailer knew a CICS region had stopped after temporary storage ran out, but not why; EVA traced it to a localized, application-driven problem and pointed to the likely source of the workload. Rocket says EVA also uses a consumption-based pricing model and can work with customers’ preferred large language model providers, while estimating a typical deployment could deliver a 3.2-times annual return on investment. That’s Rocket’s own estimate, not an independent one.
My take — AI-written commentary, not fact-checked reporting
This is the right way to do enterprise AI on mainframes: make it prove itself, keep the old controls, and resist the temptation to call every demo “autonomy.” PlanGuard sounds less glamorous than an unrestricted agent, which is exactly why it’s credible. In this market, the companies that skip the guardrails usually end up sponsoring a very expensive lesson in production.
Read more about this at: SiliconANGLE