Cloud infrastructure was built for predictable human behavior—search, click, scroll, stream. AI agents break that model. They spawn sub-agents, hammer APIs, query databases, and vanish in seconds. That mismatch is driving a major architectural shift at AWS.
This week, AWS launched the next generation of OpenSearch Serverless, a fully managed search and vector database. The headline change: complete separation of compute from storage. In the previous serverless version, you always paid for at least one idle compute node because storage and compute were coupled. Now compute can scale from zero to peak capacity in seconds, then back to zero. You only pay for what you use—like metered parking instead of a reserved garage spot.
“Agents spike without warning, go idle without notice,” said Tia White, GM for Amazon OpenSearch Service. “Enterprise search needs to keep up without paying for empty compute.”
This isn't just an AWS problem. Cloudflare reports bots already account for 31% of HTTP traffic; AI crawlers alone make up a quarter of that. Cloudflare’s Lai Yi Ohlsen predicts non-human traffic will exceed human traffic by early 2027. Google’s I/O showed users delegating purchases and bookings to agents. Enterprises are deploying internal agents at scale, generating machine-to-machine traffic that legacy infrastructure can’t handle.
The industry is responding. Databricks and Snowflake are repositioning as AI memory layers. Microsoft updated Azure for agent burst handling and shared memory. Cloudflare launched persistent environments with instant scaling last month.
OpenSearch Serverless launches with native integrations for Vercel and Kiro, letting developers deploy search backends for agents without managing infrastructure. The engineering trade-off is clear: decoupling adds complexity in consistency and data freshness, but the payoff is eliminating idle-cost overhead for spiky, unpredictable agent workloads. As agents move from prototype to production, expect more infrastructure to follow this pattern.
Source: TechCrunch