Agentic AI design patterns

An artificial intelligence agent is software that can plan, call tools, and keep working across steps, and developers are now publishing playbooks for running those systems in production. (link.springer.com) One of the most-circulated references is *Agentic Design Patterns*, a 2025 Springer book by Antonio Gullí. The table of contents shows 424 pages covering prompt chaining, routing, parallelization, reflection, tool use, planning, memory management, human-in-the-loop review, inter-agent communication, guardrails, and evaluation. (link.springer.com) (irp.cdn-website.com) Those chapters map a field that has moved beyond single chatbot prompts. The book’s foreword says developers are shifting from systems that only process information to systems that can “reason, plan, and act” on ambiguous tasks. (link.springer.com) The infrastructure pattern getting shared alongside those guides is simple: keep the agent workers disposable, and move the durable parts somewhere else. Kubernetes says stateless workloads fit interchangeable Pods, while stateful workloads need StatefulSets and persistent identity. (kubernetes.io) In that setup, Apache Kafka acts like a task conveyor belt. Kafka’s documentation says it lets systems publish and subscribe to streams of records, store them durably, and process them as they occur, which makes it a common fit for job ingestion and retries. (kafka.apache.org 1) (kafka.apache.org 2) Redis fills a different role: fast shared memory for session state, caches, and coordination data. Redis’s Kubernetes documentation pairs that speed with persistent volumes and operator-based management when teams need recovery across restarts and failures. (redis.io 1) (redis.io 2) (redis.io 3) Cost control is part of the pattern too. Kubernetes documents horizontal autoscaling for Pods, node autoscaling for cluster capacity, and event-driven scaling through Kubernetes Event Driven Autoscaler, or KEDA, so teams can add workers when queues grow and remove them when they empty. (kubernetes.io 1) (kubernetes.io 2) (kubernetes.io 3) A separate stream of projects is trying to package all of that into an “agent operating system.” MateClaw describes itself as a self-hosted artificial intelligence operating system and its public repository lists multi-agent orchestration, database-loaded agent configs, tool integration, and layered memory among its core features. (claw.mate.vip) (github.com) (gitee.com) The appeal is not that any one component is new. The shift is that developers are treating agents less like demos and more like distributed software, with design patterns for reasoning and separate patterns for queues, memory, scaling, and failure recovery. (link.springer.com) (kubernetes.io) (kafka.apache.org) (redis.io) That is why a 424-page pattern book, Kubernetes deployment recipes, and new multi-agent runtimes are landing in the same conversation. They all answer the same 2026 question: how to make an agent keep working after the prototype works once. (irp.cdn-website.com) (claw.mate.vip)