Agentic AI
Frameworks
How modern AI systems think, act, use tools, and remember — a visual guide to the frameworks powering autonomous intelligence.
LLM Backbone
The large language model that reasons, plans, and generates actions. Examples: GPT-4, Claude, Gemini, Llama.
Tools / Actions
External capabilities the agent can invoke: web search, code execution, database queries, APIs, file I/O.
Memory
Short-term (context window), long-term (vector DB), and episodic memory for recalling past interactions.
Planning
Breaking a complex goal into steps. Strategies include Chain-of-Thought, Tree of Thoughts, and MCTS.
Perception
Inputs the agent can process: text, images, audio, structured data, API responses, environment state.
Feedback Loop
The agent observes results of its actions and uses that observation to adjust the next step or retry.
Decompose goal into subtasks
Decide next action (ReAct / CoT)
Search / Code / API / File
Store observations + results
to Reasoning
LangChain is the most widely-used framework for building LLM applications. It provides composable “chains” of prompts, tools, memory, and agents. Its Agent module wraps an LLM with tools and a reasoning loop, allowing it to autonomously select which tool to use at each step.
- Chain-based composition of LLM calls
- 100+ pre-built tool integrations
- Multiple memory backends (Redis, Pinecone)
- Supports ReAct & OpenAI function calling
- LangSmith for tracing & observability
- Python & JavaScript SDKs
Example Use Case
A research agent that takes a question, searches the web with SerpAPI, reads Wikipedia pages, stores findings in memory, and synthesizes a structured report — all in one automated pipeline.
AutoGPT was one of the first autonomous agent systems to go viral. It gives an LLM a high-level goal and lets it self-direct: writing its own subtasks, executing them, storing results to files, and looping until the goal is complete. It’s a “set it and forget it” style agent.
- Fully autonomous goal-driven operation
- Built-in file read/write capabilities
- Web browsing & web scraping
- Self-generated task queue
- Plugin system for extensibility
- Docker-based sandboxed execution
Example Use Case
Given the goal “Research competitors and write a market analysis report,” AutoGPT browses the web, saves notes, generates an outline, writes sections, and produces a complete document without further human input.
CrewAI takes a multi-agent approach where multiple specialized AI agents collaborate as a “crew.” Each agent has a role, goal, and backstory. They work together sequentially or in parallel, delegating tasks and sharing context to solve complex problems no single agent could handle alone.
- Role-based multi-agent collaboration
- Sequential & hierarchical task flows
- Inter-agent delegation & communication
- Built on top of LangChain tools
- Agent memory & context sharing
- Human-in-the-loop checkpoints
Example Use Case
A startup analysis crew: a Researcher agent finds data, an Analyst agent interprets financials, a Writer agent drafts the report, and a Reviewer agent critiques and improves it — all coordinated automatically.
LangGraph models agentic workflows as a directed graph (nodes and edges) rather than simple chains. This enables complex flows with cycles, conditional branching, and parallelism. Each node is an LLM call or tool; edges define transitions based on state. It’s ideal for stateful, long-running agents.
- Graph-based stateful agent architecture
- Native support for cycles & loops
- Conditional branching between nodes
- Parallel subgraph execution
- Persistent state across interruptions
- Human-in-the-loop at any node
Example Use Case
A customer support agent that routes queries to different specialist sub-graphs (billing, technical, returns), loops back for clarification if needed, and escalates to a human when confidence drops below a threshold.
ReAct (Reasoning + Acting) is not a framework but a foundational prompting pattern. The LLM interleaves Thought (chain-of-thought reasoning), Action (tool call), and Observation (tool output) in a loop. Most major frameworks implement this pattern under the hood as their core agent loop.
- Thought → Action → Observation loop
- Synergizes reasoning with acting
- Reduces hallucination via grounding
- Works with any LLM via prompting
- Foundation of LangChain agents
- Interpretable step-by-step traces
Example Trace
Thought: “I need to find the population of India.” → Action: search(“India population 2024”) → Observation: “1.44 billion” → Thought: “Now I can answer.” → Final Answer: “India’s population is approximately 1.44 billion.”
Code Generation & Debugging
Agents write, execute, test, and debug code iteratively. They read error messages, search documentation, and fix bugs autonomously until tests pass.
DevOpsResearch & Synthesis
Agents browse the web, read papers, extract key findings, cross-reference sources, and generate comprehensive literature reviews.
ResearchEmail & Calendar Automation
Agents read emails, schedule meetings, draft replies, and manage calendars by integrating with Gmail, Outlook, and calendar APIs.
ProductivityE-Commerce Operations
Agents monitor inventory, update pricing, respond to customer queries, and process returns — running entire workflows with minimal oversight.
BusinessClinical Decision Support
Medical agents retrieve patient records, search clinical literature, cross-check drug interactions, and present evidence-based recommendations.
HealthcareFinancial Analysis
Agents pull market data, run quantitative models, analyze earnings reports, and generate investment briefs with sourced evidence.
Finance| Framework | Agent Type | Learning Curve | Multi-Agent | Best For |
|---|---|---|---|---|
| LangChain | Single / Chain | MEDIUM | PARTIAL | Tool-using pipelines, RAG apps |
| AutoGPT | Autonomous | LOW | NO | Fully autonomous open-ended tasks |
| CrewAI | Multi-Agent | MEDIUM | YES | Role-based team workflows |
| LangGraph | Stateful Graph | HIGH | YES | Complex stateful, cyclical workflows |
| ReAct | Pattern / Loop | LOW | NO | Foundation layer for any agent |