Agent Architectural Patterns

To build a comprehensive and authoritative reference on the types of AI agents, the best approach is to divide the ecosystem into two main categories:

1. Architectural Patterns — how the agent is built and orchestrated under the hood.
2. Domain-Specific Agents — what the agent practically does in production.

Regardless of vertical, the execution logic under the hood fits into one of the architectural models below. Domain specialists then layer tools, memory, and guardrails on top of those engines.

These are the logical engines. Regardless of what the agent actually does, the execution logic under the hood fits into one of these models.

Description: The foundational pattern of modern agentic workflows. The agent operates in a continuous loop of thinking about the problem, acting (calling a tool), observing the tool's output, and then deciding the next step.

Characteristics: A synchronous loop governed by a structured prompt following the Thought → Action → Observation → Final Answer framework.

Advantages: Extremely versatile, easy to build, and simple to debug. It serves as the gold standard for general tasks where the path to the solution is not known from the start.

Use cases: General virtual assistants, customer support chatbots hooked to databases, and generic analytical problem solvers.

Technologies / frameworks: LangChain (ReAct Agent), LlamaIndex, Anthropic Tool Use.

Description: Acts as a dispatcher or gateway. It takes the user's input, classifies the intent, and routes the request to the most appropriate tool, specialized sub-agent, or specific prompt.

Characteristics: Fast execution (usually single-step, no loops), focused purely on classification and parameter extraction.

Advantages: Saves tokens (by not loading unnecessary tools into the context window), reduces latency, and drastically lowers the risk of hallucinations.

Use cases: Customer service ticket triage, AI API gateways, routing between expensive models (GPT-4) and cheaper ones (GPT-4o-mini) depending on task complexity.

Technologies / frameworks: Semantic Router, OpenAI Structured Outputs, LangGraph.

Description: A central agent (the Orchestrator) receives a complex task, breaks it down into subtasks, and delegates each part to different worker agents. These workers operate in parallel or sequentially, each armed with a specific prompt and toolset.

Characteristics: High delegation, parallel execution, and strictly defined roles and personas (e.g., a researcher, a writer, a reviewer).

Advantages: Allows the system to solve absurdly complex tasks that would otherwise cause a single prompt to collapse due to instruction overload.

Use cases: Creating extensive market research reports, end-to-end software development (manager + programmer + QA), full-scale marketing campaigns.

Technologies / frameworks: CrewAI, Microsoft AutoGen, LangGraph.

Description: An iterative workflow focused on excellence. One agent (the Actor) generates a first draft of the work, and a second agent (the Critic) evaluates that draft against a set of rules, providing feedback. The Actor then refines the work based on the feedback until the Critic approves it.

Characteristics: Heavily focused on self-reflection and cyclic refinement with strict stopping criteria.

Advantages: Exponentially increases the quality of the final output. It corrects reasoning flaws or tool-calling errors before the user ever sees the response.

Use cases: Source code review and refactoring, technical and literary translation, compliance auditing in legal documents.

Technologies / frameworks: Reflexion Pattern, LangGraph, LlamaIndex Workflows.

These are the market-consolidated use cases — agents that have received implementations so specific that they have evolved into their own subcategories. Each domain specialist typically composes one or more architectural patterns from Part 1.

Description: Agents integrated directly into the user's Integrated Development Environment (IDE) or terminal. They can read entire repositories, create files, run terminal commands, test code, and push commits.

Characteristics: Deep access to the file system, secure execution in sandboxes (Docker), and intensive use of LSP (Language Server Protocol) to understand the codebase.

Advantages: Near-total autonomy in closed-scope tasks, dramatically speeding up refactoring and resolving issues reported on GitHub.

Use cases: Bug resolution, framework migration, automated test creation, automatic documentation of legacy systems.

Technologies / frameworks: Devin (Cognition), Aider, Cline (Roo Code), SWE-agent.

Description: The evolution of semantic search. Agentic RAG doesn't just do a passive search in a vector database; it formulates its own questions, evaluates if the retrieved documents contain the answer, breaks the question into smaller parts, and cross-references data from multiple sources until it finds the complete answer.

Characteristics: Query expansion, hierarchical search, and treating retrieval as an interactive tool (retrieval-as-a-tool).

Advantages: Solves complex questions where answers are fragmented across multiple different documents, overcoming the context limitations of passive RAG.

Use cases: Legal due diligence, cross-referenced medical/academic research, analyzing massive financial contracts.

Technologies / frameworks: GraphRAG, LlamaIndex (Agentic RAG pipelines), Pinecone, Milvus.

Description: Agents that control a web browser just like a human would. They read the page code (DOM) or "look" at the screen using computer vision, deciding where to click, what to type, and how to extract information from complex UIs.

Characteristics: Multimodal processing (image + text), dynamic mapping of buttons and forms (even when the UI changes).

Advantages: Can automate integrations with legacy systems or closed websites that do not offer public APIs. It is the end of fragile, traditional web scraping.

Use cases: Competitor price scraping on bot-blocked sites, automating data entry in legacy ERPs, end-to-end user interface testing.

Technologies / frameworks: Anthropic Computer Use, Browserbase, MultiOn, Skyvern, Playwright + LLMs.

Description: Unlike interactive chatbots that respond instantly, this agent is triggered by an event and can run in the background for hours or days, sending status notifications until its mission is accomplished.

Characteristics: Event-driven architecture, heavy reliance on long-term memory (so it doesn't forget what it was doing if paused), asynchronous execution.

Advantages: Decouples the agent from the user chat interface, enabling massive back-office workflows.

Use cases: Daily scraping and compilation of market news/leads, monthly consolidation of tax data from multiple sources, continuous IT security monitoring (SOC agents).

Technologies / frameworks: LangGraph (using checkpointers to pause/resume the graph), Temporal.io, Inngest.