Architecture Patterns

Agentic Architecture Patterns

A visual reference covering 8 core architectural patterns in Agentic AI and how major frameworks implement them. Explore patterns, compare frameworks, and reference this when designing your own systems.

Part 1 — Architecture Patterns

Each pattern represents a different way to structure the loop between an LLM, tools, memory, and humans. Most real systems combine two or more patterns.

Filter

Pattern 01

Simple LLM

One-shot prompt → response, no tools or persistent memory

Low complexity

Characteristics

▸ Stateless — each request is independent
▸ No external tools or API calls
▸ Fastest and cheapest to run
▸ Context window is the only “memory”

Best for

Q&ASummarizationClassificationTranslationDrafting

Pattern 02

Tool-Calling Agent (ReAct)

LLM reasons, decides which tool to call, observes the result, repeats

Medium complexity

Characteristics

▸ Thought → Action → Observation loop (ReAct pattern)
▸ Tools can be APIs, databases, code runners, search engines
▸ Runs until the LLM decides the task is complete
▸ Requires tool definitions and output parsing

Best for

Web searchCode executionAPI integrationCalculatorFile I/O

Pattern 03

RAG — Retrieval-Augmented Generation

Retrieve relevant context from a knowledge store before generating

Medium complexity

Characteristics

▸ Splits retrieval (vector search) from generation (LLM)
▸ Dramatically reduces hallucination on factual queries
▸ Knowledge is updateable without retraining the model
▸ Retriever and LLM can use different models

Best for

Document Q&AEnterprise searchComplianceCustomer supportCode search

Pattern 04

Memory-Augmented Agent

Agent reads/writes to short-term and long-term memory stores

Medium complexity

Characteristics

▸ Short-term memory: recent conversation turns (in-context)
▸ Long-term memory: vector store / database (persistent across sessions)
▸ Can personalize responses based on past interactions
▸ Memory retrieval adds latency and cost

Best for

Personalized assistantsLong-running projectsCRM botsResearch agents

Pattern 05

Planner–Executor

Separate planning from execution — one LLM makes the plan, another executes it

High complexity

Characteristics

▸ Planner decomposes goal into ordered sub-tasks
▸ Executor runs each step with access to tools
▸ Separation allows specialized models for each phase
▸ Plan can be revised if a step fails (re-planning loop)

Best for

Complex task automationReport generationCode generation pipelinesResearch synthesis

Pattern 06

Multi-Agent Orchestration

A supervisor routes subtasks to specialist agents and aggregates results

High complexity

Characteristics

▸ Supervisor (orchestrator) agent manages task routing
▸ Specialist agents run in parallel with domain expertise
▸ Results are aggregated by the supervisor
▸ Each agent can have its own tools and memory

Best for

Parallel researchSoftware development teamsContent pipelinesData analysis

Pattern 07

Event-Driven / Async Agent

Agent is triggered by external events — webhooks, schedules, messages

High complexity

Characteristics

▸ Agents wake up in response to triggers, not user prompts
▸ Message queues decouple event producers from consumers
▸ Can run multiple agent instances concurrently
▸ Suitable for long-running background automation

Best for

Workflow automationMonitoring & alertsETL pipelinesScheduled reportsWebhooks

Pattern 08

Human-in-the-Loop

Agent pauses at checkpoints for human review before continuing

Medium complexity

Characteristics

▸ Human review gate before high-stakes actions
▸ Agent can present options, human selects or corrects
▸ State is persisted while waiting for human input
▸ Reduces risk in agentic systems with real-world side effects

Best for

Content moderationFinancial decisionsMedical workflowsLegal reviewDeployment gates

Pattern Comparison

Pattern	Complexity	Memory	Tools	Parallel	Best For
Simple LLM	Low	✗	✗	✗	Q&A, summarization, classification
Tool-Calling / ReAct	Medium	✗	✓	✗	API calls, calculations, web search
RAG	Medium	External	✓	✗	Document Q&A, knowledge retrieval
Memory-Augmented	Medium	✓	✓	✗	Long conversations, personalization
Planner–Executor	High	Optional	✓	✗	Multi-step task decomposition
Multi-Agent Orchestration	High	Per-agent	✓	✓	Parallel research, complex pipelines
Event-Driven	High	Optional	✓	✓	Automation triggers, async workflows
Human-in-the-Loop	Medium	Optional	✓	✗	High-stakes decisions, content review

Part 2 — Framework Landscape

The major open-source frameworks in 2025–26, each implementing one or more of the patterns above. Most teams pick one primary framework and extend it.

LangGraph

Directed graph topology with time-travel debugging and the lowest latency in benchmarks. 400+ companies in production.

Graph-based stateful agentsOpen Source

✓ Pros

Graph visualization + checkpointing for debugging
Production-grade: LinkedIn, Uber, Replit, Elastic

✗ Cons

Steeper learning curve than simpler frameworks
Tied to LangChain ecosystem

Notable users

LinkedInUberReplitElasticAppFolio

OpenAI Agents SDK (ex-Swarms)

The simplest explicit handoff model. Stateless by design; provider-agnostic (100+ LLMs). Prioritizes clarity over abstraction.

Multi-agent handoffsOpen Source

✓ Pros

Minimal abstraction — easy to reason about
Provider-agnostic, works with any OpenAI-compatible API

✗ Cons

No built-in state persistence between runs
Newer (Mar 2025); smaller production footprint

Notable users

OpenAI ecosystemExperimental / prototyping

Google ADK (Agent Development Kit)

Unique A2A (Agent-to-Agent) protocol lets ADK agents invoke LangGraph, CrewAI, or other framework agents. Native multimodal (images, audio, video).

Event-driven hierarchical treeOpen Source

✓ Pros

A2A protocol: cross-framework interoperability
Native multimodal support via Gemini

✗ Cons

Newest (Apr 2025) — smallest production footprint
Best on Google Cloud / Vertex AI

Notable users

Google CloudVertex AI workloads

CrewAI

Role abstractions (Manager, Researcher, Writer) make team-based workflows easy to design. Now the default runtime for LangChain agents.

Role-based multi-agentOpen Source + Commercial

✓ Pros

Intuitive role metaphor — easy for non-engineers to design
LLM-agnostic: assign different models by cost/latency per role

✗ Cons

Manager agent adds ~3× token cost for simple single-tool calls
Limited checkpointing vs LangGraph

Notable users

LangChain (default runtime)Business automation teams

AutoGen / Microsoft Agent Framework

Actor model enables group-chat and debate patterns. AutoGen merged with Semantic Kernel into Microsoft Agent Framework (Oct 2025) for production use.

Conversational multi-agent (actor model)Open Source

✓ Pros

Cross-language: Python + .NET support
Built-in OpenTelemetry observability

✗ Cons

AutoGen in maintenance mode since Oct 2025 — migrate to MS Agent Framework
Higher cost per query vs CrewAI / LangGraph

Notable users

Microsoft ecosystemEnterprise .NET teams

LlamaIndex Agents

Purpose-built for RAG + agentic retrieval. AgentWorkflow (2025) adds streaming and orchestration on top of best-in-class document indexing.

Agentic RAGOpen Source

✓ Pros

Best-in-class RAG pipeline (hybrid retrieval, self-correction)
35% retrieval accuracy improvement in 2025 benchmarks

✗ Cons

Less suited for general orchestration tasks
Higher setup effort for non-document use cases

Notable users

AWS (Bedrock examples)Document-heavy enterprises

Haystack

Modular DAG architecture — every node is independently testable and replaceable. Strong enterprise production track record.

DAG pipeline agentsOpen Source + Commercial

✓ Pros

Explicit control: no black-box automation
Production-proven at Airbus, NVIDIA, European Commission

✗ Cons

More verbose setup vs higher-level frameworks
Steeper learning curve for complex agentic patterns

Notable users

AirbusNVIDIAThe EconomistEuropean CommissionLufthansa

Semantic Kernel

Best for .NET/Microsoft teams building plugin-based agent pipelines. Merging into Microsoft Agent Framework (Oct 2025) as the plugin layer.

Plugin-based orchestrationOpen Source

✓ Pros

Best .NET / C# support of any framework
Rich plugin ecosystem for Microsoft 365, Azure services

✗ Cons

Transitioning into MS Agent Framework — use that for new projects
Python support lags behind .NET

Notable users

Microsoft 365 integrations.NET enterprise teams

LangGraph vs OpenAI Agents SDK vs Google ADK

The three most-discussed graph/multi-agent frameworks in 2025–26

Dimension	LangGraph	OpenAI Agents SDK	Google ADK
Topology	Directed graphs with conditional edges	Explicit agent-to-agent handoffs	Hierarchical agent tree + A2A protocol
State	Checkpointed (reducer-driven, time-travel debug)	Stateless — context variables only	Session-based, persistent
Multimodal	Text-based	Text-based	✓ Native (images, audio, video)
Interop	LangChain ecosystem native	Provider-agnostic (100+ LLMs)	A2A protocol — invokes other frameworks
Deployment	LangGraph Platform (hosted)	Client-side (minimal infra)	Vertex AI + Google Cloud
Released	2023 (GA)	Mar 2025	Apr 2025
Users	LinkedIn, Uber, Replit, Elastic	Production upgrade of Swarms	Google Cloud ecosystem

Framework × Pattern Matrix

●Primary ◐Supported ○ Not typical

Framework	Simple LLM	Tool-Calling	RAG	Memory	Planner–Exec	Multi-Agent	Event-Driven	Human Loop
LangGraph	○	●	◐	●	●	●	◐	●
OpenAI Agents SDK	○	●	◐	○	◐	●	◐	◐
Google ADK	○	●	◐	◐	◐	●	●	◐
CrewAI	○	●	◐	◐	◐	●	○	◐
AutoGen / MS Agent	○	◐	◐	◐	●	●	◐	●
LlamaIndex Agents	○	●	●	●	◐	◐	◐	○
Haystack	○	●	●	◐	●	◐	◐	◐
Semantic Kernel	○	●	◐	◐	●	●	◐	◐

📊

2026 Performance Benchmarks

Cost per query

CrewAI — $0.12–0.15

LangChain — $0.18

AutoGen — $0.35

Latency ranking

🥇 LangGraph (lowest)

🥈 OpenAI Agents SDK

🥉 CrewAI

Production readiness

✓ LangGraph — GA

✓ MS Agent Framework — 1.0 (Apr 2026)

✓ Haystack Enterprise

✓ CrewAI — v1.0 (late 2025)

Benchmarks sourced from public comparisons as of April 2026. Costs vary by model and task complexity.

← Back to courses