Rocket Software adds governed AI agents to mainframe platform EVA
Rocket Software has announced an expanded version of EVA, its agentic AI platform for mainframe environments, adding capabilities that allow AI agents to reason across operational data and automate tasks within policy-governed boundaries. The Waltham, Massachusetts-based company, which is privately held and backed by Bain Capital Private Equity, says the platform is currently in pilot with organisations across financial services, government, insurance, retail and telecoms.
The headline addition is Rocket PlanGuard, a security layer that inserts a policy decision point between AI reasoning and system execution. PlanGuard provides identity controls and just-in-time, contextually scoped authorisation, with all agent activity logged for audit purposes. The company describes the policy architecture as patent-pending and positions it as a prerequisite for deploying autonomous AI agents safely in environments where batch failures or misconfigured access can have immediate downstream consequences.
Addressing the mainframe skills gap
A Hanover Research study commissioned by Rocket Software found that 81% of financial services IT leaders identify a mainframe skills gap as very or extremely significant, and 87% believe AI will help narrow it within two years. A separate Rocket survey found 94% of the same cohort rank enhancing IT operations with AI as a high or top priority.
Those figures reflect a structural problem that has been building for a decade: the cohort of engineers who grew up on IBM z/OS is retiring faster than it can be replaced, and the tooling to onboard junior staff has historically been opaque and manual-intensive. EVA attempts to address this by surfacing operational insights through natural language queries, capturing institutional knowledge and reducing reliance on specialists for tasks such as job failure pattern analysis, IBM CICS queue diagnostics, and IBM Db2 performance tuning.
Milan Shetti, president and chief executive of Rocket Software, said the company is "helping customers apply agentic AI to mission-critical systems with speed, confidence, and control, closing the skills gap and putting that expertise within reach of every enterprise team."
Competitive context and regulatory implications
The mainframe AI tooling market is a specialist niche, but it is attracting renewed attention as enterprises reassess the cost of migrating off z/OS against the cost of modernising in place. Rocket's direct competitors in this space include Broadcom, which absorbed CA Technologies' mainframe portfolio and has its own AI-assisted operations tooling, as well as IBM itself, which offers Watson-based observability and AIOps capabilities natively on the platform. Smaller vendors and systems integrators are also building LLM-connected tooling on top of open APIs exposed by z/OS.
Rocket's model-agnostic architecture is a differentiator worth watching: by not locking EVA to a single foundation model, the company gives enterprise buyers the flexibility to satisfy data-residency and sovereignty requirements by swapping in on-premises or region-specific models. That matters increasingly under GDPR, the EU AI Act's general-purpose AI provisions, and emerging financial-services AI governance frameworks such as the UK FCA's model-risk guidance.
The platform's inclusion of post-quantum cryptography risk identification as a security use case is also notable. NIST finalised its first PQC standards in 2024, and regulated industries with long data-retention obligations, particularly banking and insurance, are beginning to score cryptographic risk across legacy estate. An AI agent that can surface PQC exposure in mainframe configurations gives compliance teams a head start on migration planning.
Rocket says customers can move from installation to actionable insights within days. It has not disclosed specific performance benchmarks, per-agent pricing, or named any reference customers from the current pilot programme. The company's next credibility milestone will be converting pilot participants into publicly referenceable case studies with quantified operational outcomes.