Agentic System Patterns

Most people think an artificial intelligence agent is one smart model with a long prompt. Google’s own architecture guides say the real choice starts with system design: task complexity, latency, cost, and whether one agent or several agents should handle the work. (docs.cloud.google.com) That is why these “agentic patterns” keep showing up. They are not magic model tricks; they are reusable ways to split work so a language model does one step, passes state forward, or hands the job to a better specialist. (docs.cloud.google.com) The simplest pattern is prompt chaining. Google’s Agent Development Kit describes it like an assembly line: one agent parses a document, the next extracts structure, and the next writes the summary. (developers.googleblog.com) That sounds basic, but it fixes a real failure mode. When one giant prompt tries to read, decide, format, and verify in a single shot, errors compound and debugging turns into searching one huge blob of instructions for the bad line. (developers.googleblog.com) Routing is the next pattern. Instead of sending every request to the same agent, a coordinator examines the input and dispatches it to the right specialist, the way a hospital triage desk sends a broken bone to one room and chest pain to another. (developers.googleblog.com) Parallelization takes work that does not depend on a single sequence and runs pieces at the same time. Google’s guidance frames this as a way to trade extra compute for lower latency, which matters when users expect interactive replies instead of waiting for a long serial pipeline. (docs.cloud.google.com) Reflection is the pattern where one model pass checks another model pass before anything is returned or executed. In practice, that means adding a critic or evaluator loop so the system can catch bad reasoning, weak formatting, or missing constraints before the answer leaves the box. (developers.googleblog.com) Tool use is where the conversation turns into software. Google’s March 2026 protocol guide says the Model Context Protocol gives agents a standard way to discover and call tools, so developers do not have to hand-write a separate custom integration for every database, file system, or application programming interface. (developers.googleblog.com) Once several agents need to cooperate, the problem starts looking less like prompt writing and more like distributed systems. Google’s multi-agent architecture uses a coordinator agent, specialized subagents, and explicit communication protocols so different services can work together without being fused into one brittle super-agent. (docs.cloud.google.com) That is the part engineers immediately recognize. Routing, failure isolation, shared state, observability, and interface contracts are the same questions teams already solve in microservices, except now the workers are probabilistic models that can improvise, stall, or hallucinate. (developers.googleblog.com, docs.cloud.google.com) Google has been turning those patterns into product scaffolding. Its Agent Development Kit was introduced at Google Cloud Next 2025 as an open-source framework for building multi-agent applications, and Vertex AI Agent Builder pitches the enterprise layer for building, scaling, and governing those systems in production. (developers.googleblog.com, cloud.google.com) So the story here is not that someone found six clever prompt recipes. It is that large language model agents are being treated less like chatbots and more like software systems with pipelines, dispatchers, protocols, and guardrails, which is exactly what happens when a prototype meets real traffic. (docs.cloud.google.com, developers.googleblog.com)