Agent Design Patterns: Google and Anthropic, Side by Side

#ai #agents #llm #architecture

Short version: Agent design patterns are reusable ways to structure how a language model plans, delegates, and checks its own work. Anthropic and Google both published official guides, and they mostly agree. The real skill is not memorizing patterns, it is choosing the simplest one that solves your task, then adding structure only when it pays off.

What are agent design patterns?

An agent design pattern is a common architectural approach for organizing an agentic system: how the model connects to tools, and how one or more agents are orchestrated to finish a task (Google Cloud). Anthropic frames the same idea as "simple, composable patterns" that beat complex frameworks in practice (Anthropic).

The single most useful distinction comes from Anthropic. Workflows are systems where the control flow is fixed in code. Agents are systems where the model decides the control flow at runtime. That one line predicts a pattern's cost, latency, and how hard it is to debug.

How the two official guides line up

Google's guidance lives in two places: a Developer Blog post on eight multi-agent patterns in the Agent Development Kit (ADK) (Google Developers), and a Cloud Architecture Center document that adds trade-offs and a decision table (Google Cloud). Anthropic's guidance is a single engineering post, "Building Effective Agents" (Anthropic). Here is how the names map.

What you want to do	Anthropic	Google
Start from one capable unit	Augmented LLM	Single-agent system
Run fixed steps in order	Prompt chaining	Sequential pipeline
Send each input to a specialist	Routing	Coordinator / dispatcher
Do independent work at once	Parallelization (sectioning, voting)	Parallel fan-out and gather
Break a task down at runtime	Orchestrator-workers	Hierarchical task decomposition
Critique and improve output	Evaluator-optimizer	Generator-critic, iterative refinement
Let the model own the loop	Autonomous agents	ReAct
Pause for a person	Human oversight, checkpoints	Human-in-the-loop
Let peers collaborate freely	(not named)	Swarm
Combine several patterns	Combining and customizing	Composite / custom logic

The overlap is not a coincidence. Both teams watched the same production systems and named the same shapes.

The one rule both guides repeat

Start simple. Anthropic says to find the simplest solution possible and add complexity only when it demonstrably improves outcomes. Google says to begin with a single agent, refine its prompt and tools, and reach for multi-agent designs only when one agent starts to struggle. The reason is money and reliability: multi-agent systems can use many more model calls, and every extra model call adds latency, cost, and a new place for errors to compound.

A useful test before you add a pattern: can this task run as a single model call with good retrieval and a few examples? If yes, you do not need an agent at all (Google Cloud).

The patterns, one article each

Each link below is a short, illustrated guide with an animated diagram, when to use it, when not to, and the known failure modes.

The Augmented LLM: the foundation, one model with retrieval, tools, and memory.
Prompt Chaining: fixed steps, each feeding the next.
Routing: classify an input, then send it to a specialist.
Parallelization: run subtasks at the same time, then merge.
Orchestrator-Workers: decide the subtasks at runtime and delegate.
Evaluator-Optimizer: generate, critique, and refine in a loop.
Autonomous Agents: the ReAct loop, where the model owns control.
Human-in-the-Loop: a person authorizes high-stakes steps.
Swarm: peer agents that debate and converge with no central boss.
Composite Patterns: how real systems chain patterns together.

How to choose, in practice

Group the decision by the shape of your task (Google Cloud). If the steps are known and fixed, you are in workflow territory: sequential, parallel, or a refinement loop. If the model must plan and decide at runtime, you need dynamic orchestration: a single agent, a coordinator, hierarchical decomposition, or a swarm. If quality comes from cycles of feedback, use ReAct, a loop, or a generator-critic. If the task carries real risk, add a human-in-the-loop checkpoint regardless of the base pattern.

FAQ

Are Google and Anthropic's agent patterns different?
Mostly no. The names differ, but the shapes match: chaining equals a sequential pipeline, routing equals a coordinator, and so on. Google adds a few multi-agent patterns Anthropic does not name, such as swarm.

What is the difference between a workflow and an agent?
In a workflow the control flow is written in code, so it is predictable. In an agent the model decides the control flow at runtime, so it is flexible but costs more and is harder to debug (Anthropic).

Do I need a framework to use these patterns?
No. Anthropic recommends starting with direct API calls, since many patterns take only a few lines of code. Frameworks like Google ADK help once you standardize on multi-agent orchestration.

Which pattern should I start with?
A single augmented LLM. Add another pattern only when that one agent measurably falls short.