September 2026 Summaries

As AI adoption expands, architects must decide which workflow steps genuinely require real-time streaming and which are better handled through batch processing, based primarily on the cost of stale information rather than whether a task uses an LLM or is “agentic.” Streaming is most valuable when fresh events can alter an immediate decision, delay creates business or safety risks, or users, devices, and teams need live coordination; batch suits deterministic, replayable, consistency-sensitive, and cost-intensive work such as reconciliation, training, reporting, and verification. The recommended production approach is usually hybrid: retain canonical business state in databases or durable event logs, stream small versioned notifications or state changes to affected participants, and use batch jobs for heavy transformations, audit, and recovery. Teams should assess each workflow step using factors including latency tolerance, staleness cost, fan-out, determinism, inference expense, consistency requirements, and operational readiness, then validate choices through SLOs, failure tests, cost measurement, and domain expertise. The discussion emphasizes idempotency, event versions, durable audit trails, scoped security, selective model invocation, and planned degraded behavior, while positioning PubNub for low-latency distribution and presence and Kafka or similar systems for durable replayable event processing.