Agentic AI Borrowed Three Patterns from Microservices — and Made Each Worse

The AI industry is not inventing new distributed systems architecture. It is repeating the microservices transition — faster, and without the benefit of having already made these mistakes once.

Context

When the cloud-native movement drove adoption of microservices from roughly 2014 to 2020, engineering teams worked through a progression of hard lessons: decompose the monolith, replace synchronous calls with event-driven messaging, externalize state. Each lesson came with a period of production incidents before the pattern stabilized. Cornelia Davis at Temporal argues that agentic AI is recapitulating this same sequence — borrowing the same architectural patterns — but compressing what took years in the microservices era into months. The compressive speed is both the promise and the danger.

The Detail

The first pattern is decomposition. Microservices broke monolithic applications into independently deployed services with bounded ownership. Agentic AI is doing the same with LLMs: rather than one general-purpose agent handling every task, the emerging pattern builds specialized microagents — a transport agent, a lodging agent, an entertainment agent — orchestrated by a higher-level coordinator. The benefit is the same as in microservices: tighter scope improves reasoning performance, just as smaller services improve maintainability. Context windows constrained to specific tasks produce better LLM outputs than unbounded general agents.

The second pattern is the move from request/response to event-driven architecture. Early microservices relied on synchronous HTTP calls; production failures drove adoption of Kafka and message queues for resilience. Agentic workflows face the same ceiling: agent tasks spanning minutes or hours cannot rely on synchronous patterns. The Model Context Protocol introduced asynchronous durable state machines in November 2025, enabling tools to function across infrastructure failures — the AI equivalent of the message bus transition that microservices teams learned the hard way.

The third pattern is memory and state management. Microservices initially implemented stateless services that externalized everything to databases, placing heavy state reconstruction burden on developers. Agentic systems face a structurally identical problem: conversation history, intermediate decisions, and tool outputs must survive infrastructure restarts and partial failures. Event sourcing — recording every event in sequence and caching side-effect results — addresses this natively. Temporal’s implementation prevents re-expenditure of tokens on repeated LLM calls and enables complete state reconstruction regardless of nondeterminism in agent execution.

What makes Temporal’s positioning distinct is the abstraction layer it provides over this complexity. Developers write synchronous request/response code; the runtime delivers event-driven distributed behavior underneath. The Temporal integrations with OpenAI Agents SDK, Pydantic AI, Vercel AI SDK, and Google ADK are designed to preserve standard developer ergonomics while supplying production-grade durability — the same separation of concerns that mature microservices platforms eventually provided. This mirrors the agentic engineering discipline that platforms like Temporal are now formalizing as a distinct practice area.

Source: Temporal

Content created with AI assistance and reviewed for accuracy.