← Field Guide  ·  Tier 1 · Foundational Industry & Research

Google Cloud

Choose a Design Pattern for Your Agentic AI System (Oct 2025)

Foundational · cloud doc12 patterns
In Plain English

This page shows every pattern that Google Cloud has published or written about — 12 in total. Each one includes a simple picture of how it works, a short explanation, an example of when to use it, and its trade-offs (the good parts and the not-so-good parts).

Reading patterns from one source at a time matters because it lets you see how a single company or team thinks about building AI agents, and compare their ideas side by side. If a pattern here also appears in the bigger Deep Catalogue, you will see a link so you can read the full, detailed version.

An open reference book with a pencil in the fold under warm lamp light

1B.01Single-Agent System

SINGLE AGENT User Agentmodel + prompt Toolsdefined set Response

One agent with a defined toolset and system prompt handles a user request end-to-end. The model's reasoning interprets intent, plans steps, and selects tools.

Use caseCustomer support lookups, research assistants, early prototypes before scaling complexity.
Trade-offsSimplest to build and debug · Degrades as tool count and task complexity grow.

1B.02Multi-Agent Sequential

SEQUENTIAL Agent 1 Agent 2 Agent 3 Output

Predefined linear pipeline; output of agent N is input to agent N+1. No model orchestration — the sequence is hardcoded.

Use caseData pipelines (extract → clean → load), structured repeatable processes.
Trade-offsLow latency, low cost · Rigid; cannot skip steps.

1B.03Multi-Agent Parallel

PARALLEL Fan-out Agent A Agent B Agent CSynthesiser gathers outputs into one response.

Multiple specialists run concurrently on parts of the same task; results synthesised by a final agent.

Use caseCustomer feedback analysis (sentiment + keywords + urgency), multi-source data gathering.
Trade-offsReduced wall-clock latency · Increased resource use; synthesis logic complexity.

1B.04Multi-Agent Loop

LOOP Loop Controllerexit check Agent 1 Agent 2 Exit? Restartrunruncheckloop

Repeatedly executes a sequence of sub-agents until an exit condition is met (max iterations, custom state, quality threshold).

Use caseMonitoring/polling tasks, iterative refinement, self-correcting workflows.
Trade-offsEnables complex iterative work · Risk of infinite loop if exit condition ill-defined.

1B.05Review & Critique

REVIEW & CRITIQUE Generatorproduces Output Criticevaluates Approve Or Revise

Generator agent produces; critic agent evaluates against criteria; returns for revision or approves. An implementation of the loop pattern.

Use caseCode generation with security audit, content requiring strict constraints, legal/compliance drafts.
Trade-offsBetter output quality and safety · Additional model calls per iteration.

See also: 1A.07 Evaluator-Optimiser

1B.06Iterative Refinement

ITERATIVE REFINEMENT Refinermodify state Draft v1 Assess Draft v2 Cap hit?startassessrefineexit?

One or more agents progressively modify a result stored in session state, cycle after cycle, until a quality threshold or iteration cap.

Use caseLong-form drafting, multi-part planning, complex code development.
Trade-offsProduces polished output beyond single-shot · Latency and cost accumulate per cycle.

1B.07Coordinator

COORDINATOR Coordinator LLMdecides routing Specialist A Specialist B Specialist C

A central LLM-orchestrated agent analyses requests, decomposes them, and dynamically routes to specialist agents. Differs from Parallel by using a model to route rather than hardcoded dispatch.

Use caseCustomer service routing (order / return / refund), structured business processes needing adaptive routing.
Trade-offsFlexible routing · Higher token throughput and cost than hardcoded dispatch.

1B.08Hierarchical Task Decomposition

HIERARCHICAL DECOMPOSITION Root Agent Sub-Supervisor A worker ×2 Sub-Supervisor B worker ×3 Sub-Supervisor C worker ×1

Multi-level hierarchy: root agent decomposes goal; sub-agents further decompose; leaves execute. Implementation of the Coordinator pattern at depth.

Use caseOpen-ended research projects, planning + synthesis, complex multi-domain problems.
Trade-offsHandles ambiguity and scale · Architectural complexity; many model calls.

Cross-ref: Deep Catalogue PLN-PLN-01, ORC-PLN-03 →

1B.09Swarm

SWARM Dispatcherroutes initial Peer Aall-to-all Peer Ball-to-all Peer Call-to-all

All-to-all communication among specialists; a dispatcher routes the initial request but does not orchestrate. Agents hand off to each other until exit condition.

Use caseOpen-ended product design, multi-expert debate, creative problem-solving.
Trade-offsRich emergent collaboration · Most complex pattern; risk of non-convergence.

Cross-ref: Deep Catalogue ORC-ACT-02 →

1B.10ReAct (Reason + Act)

ReAct Thoughtreason Actiontool call Observationresult Thoughtnext? Answer

Iterative loop of Thought → Action → Observation. Agent reasons in natural language, acts via tools, observes results, continues until done or capped.

Use caseComplex dynamic tasks needing continuous planning — robotics, troubleshooting, exploratory research.
Trade-offsTransparent reasoning trace · Higher end-to-end latency; error propagation across steps.

Cross-ref: Deep Catalogue REA-PLN-01 →

1B.11Human-in-the-Loop

HUMAN-IN-THE-LOOP Agent Work Checkpoint Human Reviewapprove ·modify · Continue

Integrates explicit checkpoints where the agent pauses for a human to approve, correct, or provide input before continuing.

Use caseHigh-stakes financial decisions, sensitive data release, subjective creative approval, compliance validation.
Trade-offsAdds safety and accountability · Adds latency; requires external approval UI.

Cross-ref: Deep Catalogue HIL-DEC-01, HIL-END-01 →

1B.12Custom Logic

CUSTOM LOGIC Input Parallel Check Conditionalif/else Path A or B Output

Maximum flexibility: bespoke orchestration code mixing conditionals, parallel runs, and sequential flows. Used when no standard pattern fits.

Use caseBranching business processes — e.g. refund flow with parallel verification then conditional routing.
Trade-offsFine-grained control · Highest development and maintenance complexity.
Also in Tier 1
Anthropic 8Antonio Gulli 17Academic Foundations 17