A generative agent produces an output, a critic agent evaluates it, and it loops back to the generator if the criteria are not met.

Used for taking unstructured text, extracting entities, and formatting the output into strict JSON or tabular format.

A unique Neurovn testing template. A single prompt runs parallel across multiple tier-1 models for direct A/B/C cost and latency comparison.

An agent performs research, but must await human review (simulated via boolean condition) before taking destructive action.

A powerful ReAct (Reason + Act) loop pattern where the agent thinks, selects a tool, observes the result, and loops until the task is complete.

Separates concerns: A Planner breaks down the task, an Executor completes the steps sequentially, and a Synthesizer formats the final response.

A chain of specialized agents. Useful for complex workflows like content creation (Writer -> Editor -> SEO Specialist -> Publisher).

Generate and refine content with a review loop: writer drafts content, reviewer provides feedback, writer revises until approved.

Route support tickets to specialists: triage agent classifies the request, retrieves from knowledge base, then delegates to a specialist.

Extract structured data from documents: extractor agent parses raw input, validator agent checks and corrects the output.

Compare outputs from multiple models: send the same prompt to GPT-4o and Claude, then aggregate and compare results.

Simple classification: single agent node for label/decision output.

Cyclic research workflow: Agent calls search tool, reviews results, may loop for refinement.

Agent orchestrates multiple tools: database query and API call, then aggregates results.

A minimal RAG workflow: Start, RAG agent, Finish. Ideal for document Q&A.