Written by the AgileFever editorial team. Facts checked on November 18, 2026 against the official pages of the organizations named in this post. AgileFever sells training for some of the topics covered here, and we say so wherever it applies. Technical facts come from each project’s own documentation, its repository and PyPI, plus Anthropic’s engineering guidance.
Engineers and managers hear three names in every meeting about AI agent frameworks: LangGraph, CrewAI and AutoGen. Most comparisons stop at slogans. This post compares what each project’s documentation actually says about its architecture, state, multi-agent patterns, human approval, tracing, license and upkeep. It also follows one task through every option and ends with a decision table and a case for using no framework. AgileFever teaches this topic in a bootcamp, and we say where at the end. If you manage the people who will choose, read agentic AI training for engineering managers alongside this.
AI agent frameworks at a glance: status first
Status decides more than style when you compare AI agent frameworks. A framework nobody maintains is a risk you carry alone, so check it before you compare features.
| Project | Core model | License | Latest release read on November 18, 2026 | Maintenance status |
|---|---|---|---|---|
| LangGraph | Graph of shared state, nodes and edges | MIT | 1.2.22 on PyPI, uploaded Nov 14, 2026 | Active. Python plus a JS/TS library (LangGraph.js) |
| CrewAI | Crews of role-based agents, orchestrated by Flows | MIT | 1.15.41 on PyPI, uploaded Nov 12, 2026 | Active. Standalone Python framework |
| Microsoft Agent Framework | Agents plus graph-based or functional workflows | MIT (per its PyPI listing) | agent-framework 1.26.0 on PyPI, uploaded Nov 12, 2026 | Active. .NET, Python and Go (Go in public preview, per Microsoft Learn) |
| AutoGen | Agents talking in teams (Core and AgentChat layers) | MIT code (per its README) | autogen-agentchat 0.7.5, uploaded Sep 30, 2025 | Archived. Microsoft says no fixes, security patches or features |
Sources: each project’s documentation and repository page, and PyPI release data, all read on November 18, 2026. Release numbers move often; check them again before you commit.
GitHub’s own archived flag on the AutoGen repository read unset on the day we checked, while the README banner and docs say archived. Follow the maintainer’s text: treat AutoGen as legacy for any new build.
Three ways an agent run is shaped
Before features, look at what drives execution in each of these AI agent frameworks. This shape decides how you debug, test and explain the system to a reviewer.
Illustrative diagram of one refund-request task in three styles, not output from a real run.
LangGraph runs the graph you define. CrewAI hands a task to roles you name, and its docs advise starting production work with a Flow and using Crews inside Flow steps. AutoGen’s AgentChat layer lets agents exchange messages inside a team, with team types such as round robin, selector, Swarm and GraphFlow. Microsoft Agent Framework sits closest to LangGraph here: its workflows connect agents and functions through explicit paths.
What each one says about state, humans and tracing
| Area | LangGraph | CrewAI | Microsoft Agent Framework | AutoGen (legacy) |
|---|---|---|---|---|
| State and memory | Checkpointers save graph state per thread. A separate store holds long-term memory across threads. The in-memory saver loses data on restart | One Memory class. Default storage is LanceDB under a local folder; the default embedder needs an OpenAI key | Agent sessions for state, context providers for memory, durable workflow state | Team state can be saved and kept between runs |
| Multi-agent patterns | Subagents, handoffs, skills, router and custom workflows, per the LangChain docs | Sequential and hierarchical processes; hierarchical needs a manager model or agent | Sequential, concurrent, handoff, group chat and Magentic orchestrations | Round robin, selector, Swarm and GraphFlow teams |
| Human in the loop | interrupt() pauses, state is saved, then Command(resume=…) continues on the same thread ID | @human_feedback in Flows (version 1.8.0 and later). Webhook-based review is labelled Enterprise | Request-and-response steps pause a workflow for outside input | UserProxyAgent blocks the team until the person answers. max_turns pauses between runs |
| Tracing | LangSmith, LangChain’s platform for tracing and evaluation | Docs list third-party tools such as Langfuse, MLflow and Arize Phoenix. A paid platform adds monitoring | OpenTelemetry traces, logs and metrics using the GenAI semantic conventions | Not covered by active docs; the project is archived |
| License | MIT | MIT | MIT per PyPI | MIT code |
Source: official documentation pages for each project, read on November 18, 2026. A blank spot in a cell means the page did not state it, not that the feature is missing.
Two details change real designs. LangGraph resumes a paused node from its start, so any side effect before the pause, such as sending an email, must be safe to run twice. CrewAI’s default memory setup wants an OpenAI key for embeddings unless you configure another embedder, which matters if your data cannot leave your cloud.
One task, five ways, with no code
The task, an illustrative example: a customer writes in asking for a refund. The system must read the message, look up the order, check the refund policy, draft a reply, and ask a person to approve any refund above a limit.
| Approach | How you would build it | Where the human approval sits | What you debug |
|---|---|---|---|
| LangGraph | A state object carries message, order, policy result and draft. Nodes for classify, look up, check policy and draft. A conditional edge sends large refunds to an approval node | An interrupt in the approval node. State is saved, a person answers, the run resumes on the same thread | Which node ran, and the state at each step |
| CrewAI | A Flow routes the request. A crew with an order analyst, a policy reader and a writer handles the thinking step | A human feedback step in the Flow before the reply goes out | Which role produced which output, through your tracing tool |
| Microsoft Agent Framework | A workflow of executors: a triage agent, an order-lookup function, a writer agent | A request step pauses the workflow. Checkpoints let it resume | Spans per agent call and tool call in OpenTelemetry |
| AutoGen | A team of planner, order agent, writer and a user proxy exchanges messages | The user proxy blocks the team until a person replies | The message log. Treat as legacy |
| No framework | One model call with two tools, lookup_order and check_policy, in a loop your code owns | Your code checks the refund amount before the refund tool runs and asks a person if it is above the limit | Your own logs of each model call and tool call |
Illustrative example. It describes design choices, not measured results from any project.
Look at the last row. For this task, a plain loop is short, easy to test and easy to explain. The framework rows start to pay back when approvals must survive a server restart, when several people review work over days, or when many agents share state.
Which of these AI agent frameworks fits your need
| Your need | Start with | Why, from the docs |
|---|---|---|
| Fixed steps, approval gates, resume after a pause or crash | LangGraph or a Microsoft Agent Framework workflow | Both save progress at defined points and expose explicit execution paths |
| A team of named roles for research or content work | CrewAI, with a Flow around the crews | Roles, tasks and Crews are its core terms; docs advise starting with a Flow |
| A .NET shop, or Azure and Foundry as the platform | Microsoft Agent Framework | Supports .NET and Python, plus Go in public preview, and ships OpenTelemetry tracing |
| Existing AutoGen code | Plan a move to Microsoft Agent Framework | Microsoft calls it the successor and publishes a migration guide |
| Python and JavaScript teams sharing one model | LangGraph | It offers both a Python library and LangGraph.js |
| Tracing into a backend you already run | Microsoft Agent Framework, or any option plus an OpenTelemetry-compatible tool | Agent Framework emits OpenTelemetry data natively; check the others against your backend |
| One goal, a few tools, short runs | No framework | Anthropic and Microsoft’s docs both say to start simpler |
If yes, write the loop yourself and stop here.
If yes, favor LangGraph or an Agent Framework workflow, since saved state is their strength.
.NET or Azure points to Agent Framework. Python and JS together points to LangGraph.
If yes, try CrewAI and keep the Flow in charge of the order.
When to use no framework
Anthropic’s engineering guidance on building agents suggests starting with direct API calls, finding the simplest solution possible and adding complexity only when needed. Read the Anthropic guide on building effective agents. It was published in December 2024 and notes that tooling has changed since,. It separates workflows, where code sets the path, from agents, where the model decides its own steps, and names five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer.
Microsoft’s Agent Framework overview says the same in one sentence: if you can write a function to handle the task, do that instead of using an AI agent. The LangChain multi-agent page adds that a single agent with the right tools often matches a multi-agent design.
- One person owns the loop and it stays small. A framework adds terms your team must learn.
- You need exact control of prompts and retries. A thin loop keeps every call visible.
- The run takes seconds and nothing needs to pause. Saved state buys you nothing.
Move to one of the AI agent frameworks when you feel the pain it removes: approvals that outlive a process, state you keep rebuilding, or several agents that need shared rules. Related reading on tool access: Model Context Protocol explained.
A one-day trial before you commit
Build the same small task in your two shortlisted AI agent frameworks, and the plain loop as a baseline. Then answer these ten questions with evidence, not impressions.
| Question | What good looks like |
|---|---|
| Can you stop mid-run and resume after killing the process? | Same result, no repeated side effects |
| Can a new teammate trace why step 4 ran? | A trace or log shows inputs and outputs per step |
| Where does state live? | In a database you run, not in memory or a local folder |
| What leaves your network? | Prompts, embeddings and traces go only where you approve |
| Can you pin the version? | Lock file set, release notes read, upgrade path known |
| Who maintains it? | A named organization with recent releases |
| How many new concepts did your team learn? | Fewer than the baseline loop needed, or each one earns its place |
| Can you test a step alone? | Yes, with a fixed input and a checked output |
| Where is the approval gate? | In code the model cannot skip |
| What happens on a bad tool result? | A defined retry or stop, not a silent guess |
Where AgileFever fits with AI agent frameworks
We sell training on this topic, so weigh our view accordingly. The GenAI and Agentic AI BootCamp is an 8-week live virtual program. Its curriculum has a module on agent frameworks covering CrewAI, AutoGen and LangGraph, where learners build the same agent workflow in each and compare orchestration, state and control. Other modules cover memory, tool use, multi-agent collaboration, deployment and evaluation. The page asks for basic Python and says Python is not taught in the live sessions.
As listed on November 18, 2026, the bootcamp shows $2,000 per seat. Listed weekday cohorts include November 23, 2026 to January 20, 2027 and December 7, 2026 to February 3, 2027, 7:30 to 9:30 PM ET. AgileFever adds 18% GST at checkout for buyers in India; read the rupee amount in your cart.
Additional discounts are available. Contact us to know more.
To judge any course against your needs, use the agentic AI course syllabus checklist. To choose between a short workshop and a full program, read agentic AI workshop vs bootcamp. For a wider view of the options, see the best agentic AI courses, bootcamps and training programs.
Questions people ask
Which of the AI agent frameworks is best for beginners?
A plain loop is the best start, because you see every call. Among frameworks, CrewAI has the fewest terms to learn: agents, tasks, crews and flows. LangGraph asks you to design a graph and state first, which takes longer but gives control. Choose by the task, then by team skill.
Is AutoGen still maintained?
No. Microsoft’s AutoGen documentation, read on November 18, 2026, says the project is archived and no longer actively maintained, with no bug fixes, security fixes or new features. It tells new users to start with Microsoft Agent Framework and offers a migration guide.
Are LangGraph and CrewAI free to use?
Both are released under the MIT license, per their repositories, so the code is free to use. Costs come from the model calls you make and any paid platform you add for monitoring or hosting. Read each vendor’s current terms before you budget.
Does LangGraph need LangChain?
No. The LangGraph repository says it can be used without LangChain, though it integrates with LangChain products such as LangSmith for tracing. CrewAI describes itself as a standalone Python framework with its own primitives.
What replaced AutoGen?
Microsoft Agent Framework. Microsoft Learn calls it the direct successor to AutoGen and Semantic Kernel, built by the same teams. It supports .NET and Python, adds Go in public preview, and offers graph-based workflows, checkpoints and OpenTelemetry tracing.
Do I need a framework to build an AI agent?
Not always. Anthropic’s guidance suggests starting with direct API calls and adding complexity only when it helps. A model with a few tools in a loop you own is enough for many tasks. Add a framework for durable state, human approvals or many cooperating agents.
