MongoDB has released MongoDB 9.0 and opened public previews of two Atlas services aimed at the operational demands of AI applications. Atlas Infinite introduces an architecture that separates compute from storage, while Atlas Agent Engine combines agent execution, memory, retrieval and governance in one platform.
The three launches were announced at MongoDB's Investor Day. Together, they reflect a shift from building AI demonstrations to operating agents against live enterprise data, where unpredictable traffic, persistent context, identity and auditability become infrastructure concerns.
Atlas Agent Engine combines runtime, memory and governance
Atlas Agent Engine is available in public preview. MongoDB positions it as a modular layer for deploying and governing agents without tying teams to a single model or framework.
The service includes long-term memory and retrieval powered by MongoDB's Voyage AI technology. It also records actions against a human or agent identity, applies centrally managed policies and traces model calls, tool calls and memory access. These controls are intended to help teams reconstruct what an agent did and enforce approval or cost policies around its actions.
MongoDB says the platform supports open standards including Model Context Protocol (MCP) and Agent2Agent (A2A). Customers can use the memory and governance components independently or alongside the runtime, and the company says the design is neutral across AI models and frameworks.
Public-preview pricing is consumption based. MongoDB currently lists Agent Runtime at US$0.04 per 1,000 vCPU-seconds, long-term memory storage at US$0.25 per 1,000 documents and memory retrieval at US$0.50 per 1,000 documents. Preview pricing can change before general availability.
Atlas Infinite separates compute and storage
Atlas Infinite is a new Atlas deployment option designed for workloads with sharp or unpredictable demand. Unlike the existing Atlas architecture, now called Atlas Core, Infinite separates compute from storage so each can scale independently.
MongoDB says customers can adopt Infinite without changing application code because it uses the same drivers, APIs and operational tooling as Atlas Core. The preview is initially available on AWS; support for other cloud providers and Atlas for Government is planned for a later date.
The company claims Infinite cuts scaling time by more than 96% and, in internal tests using MongoDB 9.0, delivered 189% more throughput per dollar than Atlas Core. These are vendor-supplied benchmark figures rather than independent test results, and real-world gains will depend on workload shape, configuration and pricing.
Atlas Infinite uses consumption-based pricing for compute and storage. MongoDB says customers can use Core, Infinite or both depending on workload requirements.
MongoDB 9.0 focuses on throughput, encryption and resilience
MongoDB 9.0 is generally available for Atlas, Enterprise Advanced and Community Edition. Compared with MongoDB 8.0, the company claims up to twice the throughput on large instances, up to 35% faster find-one queries, up to 30% faster update-one queries and up to 20% higher throughput for transactional workloads.
Queryable Encryption has also expanded to support prefix, suffix and substring searches. This lets applications query supported fields while their sensitive data remains encrypted, although teams will still need to assess the feature's supported query patterns and operational trade-offs for their own deployments.
A new Intelligent Workload Management capability is designed to protect overloaded clusters and keep short-running operations responsive during demand spikes. MongoDB 9.0 also adds broader OpenTelemetry support and improvements to backup, restore and recovery.
A connected stack, with previews still to prove
MongoDB's pitch is that version 9.0 provides the database foundation, Atlas Infinite handles elastic scaling, and Agent Engine supplies the context and controls needed to operate AI agents. The approach could reduce the number of separate services that engineering teams must integrate, but two of the three products remain in public preview.
That preview status matters for production planning: availability, pricing, feature limits and cloud coverage may change. TTR has not independently benchmarked MongoDB 9.0 or Atlas Infinite, and has not tested Atlas Agent Engine. Performance and efficiency figures in this article are attributed to MongoDB.
