AI agents are forcing a rethink of enterprise infrastructure
AI agent infrastructure refers to the software, data, and compute layers that allow autonomous AI agents to run complex, long-lived workflows reliably across many enterprise systems while scaling elastically and keeping costs under control. That requirement is colliding with the limits of traditional enterprise automation platforms, which were built for predictable traffic patterns and human-triggered tasks. Agent workloads behave differently: interactions spike and stall, rely on vector search, and require persistent memory and governance from day one. Instead of bolting AI features onto old stacks, large platforms are redesigning their foundations around agent-native architecture. Serverless AI workloads, task routing, semantic search, and continuous evaluation are being treated as core primitives, not add-ons. The result is a structural shift in how enterprise software is planned, priced, and operated.
AWS rebuilt OpenSearch Serverless around agent-native architecture
Amazon OpenSearch Serverless has been rebuilt to support the bursty, search-heavy patterns of AI agents. According to Tia White, general manager for OpenSearch at AWS, “about 97 percent of it has been built from the ground up by the engineers on the managed service.” The new design separates storage and compute on a proprietary storage layer and allows collections to scale all the way to zero, so customers pay nothing when workloads are idle. Auto-scaling is now about 20 times faster, aiming to avoid cold starts while handling sudden spikes from serverless AI workloads. AWS says this architecture can cut costs by up to 60 percent compared with provisioned clusters running at peak capacity, thanks to aggressive down-scaling and storage compression. The roadmap extends beyond search to agent memory, knowledge graphs, and an “agentic-first” platform with built-in evaluation and governance rather than retrofitted controls.
Asana’s StackAI deal: agents as first-class enterprise users
Asana’s acquisition of StackAI, reportedly valued at USD 75 million (approx. RM345 million), signals that work management platforms now see AI agents as first-class actors inside enterprise systems. StackAI builds tooling that lets agents perform tasks across multiple SaaS applications, turning natural-language requests into actions such as updating records, triggering workflows, or collecting data from different tools. By bringing this capability in-house, Asana is moving beyond AI-assisted text and task suggestions toward an operating model where agent-native architecture sits alongside human workflows. Instead of retrofitting AI onto existing automation rules, the goal is to let agents plan, execute, and coordinate work across CRM, ticketing, analytics, and other systems through Asana. This reflects a broader shift in enterprise automation platforms: they are evolving from workflow orchestrators around humans to orchestration hubs for both humans and autonomous agents sharing the same work graph.
From retrofits to agent-native infrastructure as table stakes
Both AWS and Asana are showing that adding AI into old stacks is not enough. AWS is discarding much of OpenSearch’s earlier serverless design because agents exposed a “Swiss Army knife” problem: one product trying to be many things without being tuned for bursty, vector-heavy workloads. Asana, meanwhile, is buying dedicated agent execution tooling instead of extending legacy integrations. Across the stack, cost efficiency and cross-system execution are becoming table-stakes requirements for enterprise AI infrastructure. Platforms must support serverless AI workloads that scale to zero, provide semantic layers and observability tailored to agent workflows, and enable governance baked into the core. In this world, the winners will not only expose APIs to large language models; they will rebuild their foundations so agents can act like reliable digital employees operating safely and efficiently across every enterprise system.