Streaming is the default temperature of agent-era data

Batch data was never really fine — it was fine for humans. A dashboard refreshed nightly worked because its consumer had patience, context, and the good sense to ask "as of when?" before acting. Agents have none of those. An agent acts now, on current state, and treats whatever it reads as current state. The moment your data's consumer stops being a human with patience, batch staleness stops being a freshness footnote and becomes an error source: the agent that offers a discount for a cart abandoned four hours ago — after the customer already bought — isn't working from old data, it's acting on a world that no longer exists — and unlike the analyst, it will act at scale, confidently, thousands of times an hour.

The reverse-ETL literature names this precisely: the open feedback loop. Warehouse-centric activation syncs modeled data outward on batch cadence, and the cart-abandonment latency chain — event to warehouse to model to sync to action — is measured in hours. Human marketing tolerated that. AI decisioning demands a closed loop: decide, act, observe, retrain within seconds, which is the entire argument behind the "Agentic CDP" critique of batch activation. The same shift shows up on the read path. An agent mid-task can't wait on a warehouse queue; the real-time OLAP category — Pinot, Druid, StarRocks — exists to answer aggregation-heavy queries sub-second over continuously arriving data, at high concurrency, for consumers that aren't internal analysts. Pinot's design target of 10ms P95 was built for user-facing analytics; agents are simply the newest user-facing consumer, and the highest-volume one.

The substrate is already moving

What makes "default" credible rather than aspirational is that the streaming substrate's economics collapsed at exactly the right moment. The 2026 movement toward diskless, object-storage-native streaming — WarpStream's zero-disk architecture, AutoMQ on S3, KIP-1150 in Kafka proper — puts the durable log on the same cheap storage as the lakehouse, eliminating cross-AZ replication costs and, in "store once, read many ways" architectures, ending the Lambda-era practice of keeping separate batch and speed layers. Meanwhile lightweight backbones like NATS JetStream demonstrate what the agent-facing tier looks like: sub-millisecond core pub/sub, exactly-once tool invocation via message deduplication, LLM token streaming, agent telemetry pipelines — event infrastructure shaped for machine consumers. The pattern the streaming community calls shift-left is the temperature change in one phrase: processing and quality enforcement move into the stream, and analytics stops waiting for the nightly landing.

The honest counter-case is that "default" is not "universal," and the streaming world itself supplies the caveat. Diskless designs make latency an explicit buffering-window trade-off — batching writes to object storage adds hundreds of milliseconds, which is fine for telemetry and wrong for fraud checks; not every stream needs to be hot, and paying for heat you don't need is its own failure mode. Genuinely batch workloads remain, too: model training, monthly reconciliation, anything where recency bias runs the other way and last quarter matters more than last second. Streaming also carries real operational weight, which is exactly why event sourcing must earn its complexity rather than be adopted on vibes. The claim is about the default, not the extremes: when a new dataset arrives, the old presumption was batch-unless-proven-urgent. Agent-era consumers invert it — current-unless-proven-tolerant. Staleness becomes something you justify, budget, and declare, because the consumer can no longer smell it.