{"id":19298,"date":"2026-06-12T11:54:32","date_gmt":"2026-06-12T06:54:32","guid":{"rendered":"https:\/\/multiqos.com\/blogs\/?p=19298"},"modified":"2026-06-12T11:57:57","modified_gmt":"2026-06-12T06:57:57","slug":"ai-agent-frameworks-comparison","status":"publish","type":"post","link":"https:\/\/multiqos.com\/blogs\/ai-agent-frameworks-comparison\/","title":{"rendered":"AI Agent Frameworks Comparison: LangChain vs LangGraph vs AutoGen vs CrewAI"},"content":{"rendered":"<p><span style=\"font-weight: 400;\">60% of Fortune 500 companies are now <\/span><a href=\"https:\/\/www.businesswire.com\/news\/home\/20251119857048\/en\/CrewAI-Strengthens-AI-Agent-Operations-Platform-With-New-Product-Global-Expansion-and-AI-Course-with-Andrew-Ng\" rel=\"nofollow noopener\" target=\"_blank\"><span style=\"font-weight: 400;\">running CrewAI<\/span><\/a><span style=\"font-weight: 400;\"> in some form of production. That same year, <\/span><a href=\"https:\/\/www.langchain.com\/state-of-agent-engineering\" rel=\"nofollow noopener\" target=\"_blank\"><span style=\"font-weight: 400;\">LangGraph powers<\/span><\/a><span style=\"font-weight: 400;\"> the majority of documented enterprise agent deployments requiring durable, stateful execution. Two frameworks. Wildly different adoption stories. Both are telling a version of the truth.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">The AI agent frameworks market has fragmented faster than most enterprises can evaluate. You have LangGraph, CrewAI, AutoGen, and LangChain Core. Each with a legitimate use case, each with real production track records, and each with a failure mode that costs you months if you pick wrong.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Many organizations often spend weeks building on the wrong framework before realizing it couldn&#8217;t handle their compliance checkpoint requirements. This leads to stalled projects, delayed outcomes, and overspending on AI.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">This guide exists to prevent that. It offers a five-dimensional comparison of <a href=\"https:\/\/multiqos.com\/blogs\/ai-agent-frameworks\/\">popular AI agent frameworks<\/a>, a decision matrix, and three real implementation architectures.<\/span><\/p>\n<h2><b>AI Agent Frameworks Comparison: An Overview<\/b><\/h2>\n<p><b><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19301\" src=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/AI-Agent-Development-Frameworks-Comparison_-An-Overview.webp\" alt=\"AI Agent Development Frameworks Comparison Overview\" width=\"2048\" height=\"1368\" srcset=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/AI-Agent-Development-Frameworks-Comparison_-An-Overview.webp 2048w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/AI-Agent-Development-Frameworks-Comparison_-An-Overview-430x287.webp 430w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/AI-Agent-Development-Frameworks-Comparison_-An-Overview-1024x684.webp 1024w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/AI-Agent-Development-Frameworks-Comparison_-An-Overview-1536x1026.webp 1536w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/AI-Agent-Development-Frameworks-Comparison_-An-Overview-150x100.webp 150w\" sizes=\"auto, (max-width: 2048px) 100vw, 2048px\" \/><\/b><\/p>\n<h3><b>1. LangGraph: The Production Default for Stateful Agents<\/b><\/h3>\n<p><a href=\"https:\/\/github.com\/langchain-ai\/langgraph\/releases\" rel=\"nofollow noopener\" target=\"_blank\"><span style=\"font-weight: 400;\">LangGraph v1.2<\/span><\/a><span style=\"font-weight: 400;\"> (released May 2026) is the reference implementation for stateful enterprise agents. Architecture follows a state machine model: nodes perform work, edges define transitions, typed states carry data between steps. The 1.0 milestone shipped in October 2025 with a no-breaking-changes commitment through 2.0.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Why it dominates production:<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Checkpointing<\/b><span style=\"font-weight: 400;\">&#8211; saves full graph state at every node. If execution crashes, workers resume exactly where they stopped. Use PostgresSaver in production; skip SQLite for any workload with real concurrency.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Human-in-the-loop (HITL)- <\/b><span style=\"font-weight: 400;\">interrupts pause execution at any node. Ideal for approval gates in financial or compliance workflows.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Time-travel debugging<\/b><span style=\"font-weight: 400;\">&#8211; rewind to any historical checkpoint. Critical for diagnosing production incidents.<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">Steepest learning curve here. Expect 80\u2013150 lines of code for a first working agent and 2\u20134 weeks of ramp time for a senior engineer.<\/span><\/p>\n<h3><b>2. CrewAI: Role-Based Multi-Agent Collaboration<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">CrewAI is the fastest path to a working multi-agent system. A three-agent crew (researcher, analyst, writer) can be operational in <\/span><a href=\"https:\/\/docs.crewai.com\/\" rel=\"nofollow noopener\" target=\"_blank\"><span style=\"font-weight: 400;\">roughly 20 lines of code<\/span><\/a><span style=\"font-weight: 400;\">; an experienced engineer can ship a working prototype in an afternoon.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">CrewAI powered nearly <\/span><a href=\"https:\/\/blog.crewai.com\/lessons-from-2-billion-agentic-workflows\/\" rel=\"nofollow noopener\" target=\"_blank\"><span style=\"font-weight: 400;\">2 billion agentic system<\/span><\/a><span style=\"font-weight: 400;\"> executions in 2025, roughly 450 million per month at peak. Enterprise adoption includes PwC, IBM, Capgemini, and NVIDIA, making it one of the most popular among the AI agent frameworks.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Cost note: CrewAI attaches each agent&#8217;s role, goal, and backstory to every model call. That&#8217;s why token consumption is 3\u20135x higher than with LangGraph for equivalent workflows. Manageable for low-to-mid volume; a real budget line item for high-frequency automation.<\/span><\/p>\n<h3><b>3. Microsoft Agent Framework (formerly AutoGen + Semantic Kernel)<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">Microsoft Agent Framework 1.0 hit GA on 3 April 2026, unifying AutoGen&#8217;s multi-agent orchestration with Semantic Kernel&#8217;s enterprise foundation into a single SDK under Microsoft.Agents.AI. AutoGen and Semantic Kernel are now in maintenance mode, bug fixes only, no new features.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">What ships at 1.0:<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Single SDK<\/b><span style=\"font-weight: 400;\">&#8211; .NET and Python parity, six provider support (Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, Ollama). One-line provider swaps.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Sandboxed code execution<\/b><span style=\"font-weight: 400;\">&#8211; the AutoGen-inherited UserProxyAgent pattern writes Python, executes in Docker, checks output, and iterates. Natural fit for data analysis pipelines.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Azure-native enterprise<\/b><span style=\"font-weight: 400;\">&#8211; Azure AI Foundry integration, Entra ID authentication, enterprise support contracts.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>MCP and A2A native<\/b><span style=\"font-weight: 400;\">&#8211; full Model Context Protocol for tools; Agent-to-Agent protocol for cross-framework collaboration.\u00a0<\/span><\/li>\n<\/ul>\n<h3><b>4. LangChain Core: The Integration Substrate<\/b><\/h3>\n<p><a href=\"https:\/\/blog.langchain.com\/langchain-langgraph-1dot0\/\" rel=\"nofollow noopener\" target=\"_blank\"><span style=\"font-weight: 400;\">LangChain Core v1.0<\/span><\/a><span style=\"font-weight: 400;\"> in 2026 functions as an integration layer, not an agent runtime. Right choice for RAG pipelines, document processing, and connecting to 1,000+ LLMs, vector databases, and APIs. For complex stateful logic, LangGraph handles the runtime; the two are designed to work together.<\/span><\/p>\n<p><a href=\"https:\/\/multiqos.com\/contact-us\/\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19306\" src=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Picking-the-wrong-framework-costs-months.-Our-AI-engineers-help-you-choose-build-and-ship-the-right-one-without-the-rewrites.webp\" alt=\"Picking the wrong framework costs months. Our AI engineers help you choose, build, and ship the right one without the rewrites\" width=\"1400\" height=\"418\" srcset=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Picking-the-wrong-framework-costs-months.-Our-AI-engineers-help-you-choose-build-and-ship-the-right-one-without-the-rewrites.webp 1400w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Picking-the-wrong-framework-costs-months.-Our-AI-engineers-help-you-choose-build-and-ship-the-right-one-without-the-rewrites-430x128.webp 430w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Picking-the-wrong-framework-costs-months.-Our-AI-engineers-help-you-choose-build-and-ship-the-right-one-without-the-rewrites-1024x306.webp 1024w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Picking-the-wrong-framework-costs-months.-Our-AI-engineers-help-you-choose-build-and-ship-the-right-one-without-the-rewrites-150x45.webp 150w\" sizes=\"auto, (max-width: 1400px) 100vw, 1400px\" \/><\/a><\/p>\n<h2><b>Quick Comparison Table: Pick Before You Read Further<\/b><\/h2>\n<table>\n<thead>\n<tr>\n<th><b>Dimension<\/b><\/th>\n<th><b>LangGraph<\/b><\/th>\n<th><b>CrewAI<\/b><\/th>\n<th><b>Microsoft Agent Framework<\/b><\/th>\n<th><b>LangChain Core<\/b><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Latest Version<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">v1.2 (May 2026)<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">OSS 1.0 GA (2026)<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">v1.0 GA (3 Apr 2026)<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">v1.0 (Oct 2025)<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Best For<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Durable, stateful, compliance-heavy workflows<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Role-based multi-agent collaboration, fast prototypes<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Azure-native code execution, M365 integration<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">RAG + integration substrate (1,000+ connectors)<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Lines of Code for First Agent<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">80\u2013150<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">20\u201360<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">40\u201380<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">30\u201360<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Ramp Time (Senior Engineer)<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">2\u20134 weeks<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">2\u20134 days<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">1\u20132 weeks<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">1 week<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Token Efficiency<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Highest<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Lowest (3\u20135x overhead vs LangGraph)<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Moderate<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">High<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Native MCP Support<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes (default at 1.0)<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Native A2A Support<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Via LangChain<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes (default at 1.0)<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">Yes<\/span><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p style=\"text-align: center;\"><b>Known Enterprise Users<\/b><\/p>\n<\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Klarna, LinkedIn, Replit, Uber, Elastic<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">PwC, IBM, Capgemini, NVIDIA<\/span><\/td>\n<td style=\"text-align: center;\"><span style=\"font-weight: 400;\">Microsoft customers via Azure AI Foundry<\/span><\/td>\n<td>\n<p style=\"text-align: center;\"><span style=\"font-weight: 400;\">JP Morgan Chase, Ube, and SAP<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2><b>Five Criteria for AI Agent Framework Comparison<\/b><\/h2>\n<p><span style=\"font-weight: 400;\">Most teams pick a framework because it has lots of GitHub stars or because someone hyped it at a conference. Then six months later, they&#8217;re tearing it all down and starting over. Here are the five questions your evaluation should actually answer.<\/span><\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19304\" src=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Five-Criteria-for-AI-Agent-Framework-Comparison.webp\" alt=\"AI Agent Framework Comparison Criteria \" width=\"2048\" height=\"1752\" srcset=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Five-Criteria-for-AI-Agent-Framework-Comparison.webp 2048w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Five-Criteria-for-AI-Agent-Framework-Comparison-386x330.webp 386w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Five-Criteria-for-AI-Agent-Framework-Comparison-1024x876.webp 1024w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Five-Criteria-for-AI-Agent-Framework-Comparison-1536x1314.webp 1536w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Five-Criteria-for-AI-Agent-Framework-Comparison-150x128.webp 150w\" sizes=\"auto, (max-width: 2048px) 100vw, 2048px\" \/><\/p>\n<h3><b>1. How Easy Is It to Learn?<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">From easiest to hardest:<\/span><\/p>\n<ul>\n<li style=\"list-style-type: none;\">\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>CrewAI<\/b><span style=\"font-weight: 400;\"> is the quickest to pick up. You can build a working prototype in 20 to 60 lines of code. The way you define agents (giving them a role and a goal) reads almost like plain English, so even non-engineers on your team can look at it and tell you when something feels off.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>AutoGen<\/b><span style=\"font-weight: 400;\"> feels natural if you&#8217;ve used chat tools before. Microsoft pushed a big update from version 0.2 to 1.0, so most of the tutorials you will find on Google or YouTube are now outdated. Expect some confusion when you Google an error.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>LangGraph<\/b><span style=\"font-weight: 400;\"> is the steepest climb. You&#8217;re looking at 80 to 150 lines just for your first agent, and the way it thinks about workflows (called a &#8220;state graph&#8221;) takes most engineers 2 to 4 weeks to fully wrap their heads around.\u00a0<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>LangChain Core <\/b><span style=\"font-weight: 400;\">has a moderate learning curve, which can take 2 to weeks for a developer to master.\u00a0\u00a0<\/span><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">If you&#8217;re onboarding junior engineers who have less knowledge about <\/span><a href=\"https:\/\/multiqos.com\/blogs\/how-to-build-ai-agents\/\"><span style=\"font-weight: 400;\">how to build an AI agent<\/span><\/a><span style=\"font-weight: 400;\"> and you need them to be productive fast, CrewAI gets you there. If you want your engineers to control every tiny detail of how the agent runs, LangGraph rewards the longer learning curve.<\/span><\/p>\n<h3><b>2. Is It Ready for a Big Company?<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">LangGraph has all of the capabilities that big companies would want: if something crashes, it helps find the root cause,\u00a0 saves progress in a Postgres database, and allows a human to intervene and approve at any moment.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">When your compliance team or auditor visits your workplace, you can provide a complete record of the AI&#8217;s activity and the time it was carried out.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Your compliance team or security auditor can see exactly what the AI did and when they visit your workplace. That&#8217;s an audit trail that is required in regulated industries, such as finance or healthcare.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">CrewAI&#8217;s enterprise version features an audit trail, role access controls, managed hosting, scheduling, and monitoring dashboards. Enterprises can further customize the model by leveraging <\/span><a href=\"https:\/\/multiqos.com\/ai-development-services\/\"><span style=\"font-weight: 400;\">AI development services<\/span><\/a><span style=\"font-weight: 400;\"> and ensuring optimal business insights.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">LangChain is also among the AI agent frameworks that are enterprise-ready and provide stability with state management and control over mission-critical applications. On the other hand, AutoGen too is highly capable for enterprise use cases with multi-agent systems, helping Fortune 500 companies automate complex workflows.<\/span><\/p>\n<h3><b>3. Can You See What the AI is Actually Doing?<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">None of these frameworks gives you complete visibility into your agents out of the box. Your infrastructure team needs to know this before you sign anything.<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>LangGraph<\/b><span style=\"font-weight: 400;\"> uses LangSmith for tracing, which became smart enough to handle looping workflows in February 2026. The catch is that LangSmith costs money, and now you&#8217;ve added a vendor you can&#8217;t easily swap out.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>CrewAI<\/b><span style=\"font-weight: 400;\"> exports data using OpenTelemetry (the standard format), so it plays nicely with tools you might already have, like Langfuse or Arize. They launched their own enterprise dashboard in March 2026.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>AutoGen<\/b><span style=\"font-weight: 400;\"> logs conversations and connects to Azure Monitor. The new 1.0 version handles events better than older versions did.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>LangChain core <\/b><span style=\"font-weight: 400;\">offers observability and visibility features that help you trace the full execution tree for your agents.<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">My advice: treat monitoring as its own decision, separate from picking a framework. Pick your monitoring tool first (Langfuse, Arize, Datadog, Honeycomb, whatever fits your stack), then check whether the framework you want plays nicely with it.<\/span><\/p>\n<h3><b>4. How Much Will This Cost You?<\/b><\/h3>\n<p><a href=\"https:\/\/multiqos.com\/blogs\/ai-agent-development-cost\/\"><span style=\"font-weight: 400;\">AI Agent development cost<\/span><\/a><span style=\"font-weight: 400;\"> varies upon the tokens your AI agent uses. From cheapest to most expensive framework on tokens: LangGraph, then LangChain Core, then CrewAI, then AutoGen.\u00a0<\/span><\/p>\n<p><b>Why does LangGraph win?\u00a0<\/b><\/p>\n<p><span style=\"font-weight: 400;\">It only sends the data agents need to share, not the full back-and-forth conversation history. CrewAI, by comparison, attaches the agent&#8217;s role, goal, and backstory to every single call.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">That&#8217;s why CrewAI burns through 3 to 5 times more tokens than LangGraph for the same task. It&#8217;s not a bug. That&#8217;s just how it&#8217;s built. Here&#8217;s what that looks like in real money.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">If you run 10,000 agent jobs every month using Claude Opus, the gap between LangGraph and CrewAI works out to $51,600 a year. Once you&#8217;re running 10,000+ agent calls a day, your CFO starts paying attention.<\/span><span style=\"font-weight: 400;\"><br \/>\n<\/span><span style=\"font-weight: 400;\"><br \/>\n<\/span><span style=\"font-weight: 400;\">LangChain core is an open-source AI agent framework and is completely free to use. However, beyond the free tier, which is 5,000 traces per month, you need to pay $39\/seat\/month.<\/span><\/p>\n<h3><b>5. How Hard is It to Deploy?<\/b><\/h3>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>LangGraph<\/b><span style=\"font-weight: 400;\"> needs Postgres running in production to save its progress. You can pay for the LangGraph Platform to handle hosting for you, or run it yourself. Scaling out means multiple servers sharing one Postgres database.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>CrewAI<\/b><span style=\"font-weight: 400;\"> is the easiest to deploy on the open-source side. You just need a Python service running. The enterprise tier gives you managed hosting and scheduling, so your team barely has to think about infrastructure.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>AutoGen<\/b><span style=\"font-weight: 400;\"> requires Docker because the agents run code. If you&#8217;re already using Azure AI Services, deployment is straightforward. Outside of Azure, plan time for setting up container orchestration.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>LangChain core<\/b><span style=\"font-weight: 400;\"> is not hard to deploy because it is basically a Python library and not a standalone server. Enterprises can deploy it exactly like deploying Python code. However, if you are to deploy real-time token streaming or conversational chat interfaces, complexity increases.\u00a0<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">And one more time, for all four: multi-tenant isolation is something you&#8217;ll build yourself. Doesn&#8217;t matter which framework you pick.<\/span><\/p>\n<h2><b>Decision Matrix: When to Use Which Framework?<\/b><\/h2>\n<p><b>This is the section you show your team before the architecture meeting.<\/b><\/p>\n<p><b>Use LangGraph when:<\/b><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Workflow needs cycles, conditional branching, and explicit retry logic<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Human approval required at specific workflow steps (HITL checkpoints)<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Durable execution: agents must survive crashes and resume without data loss<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Full audit trail required for compliance or regulatory review<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Already running LangChain elsewhere in the stack<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Complex multi-step use cases: code review pipelines, financial analysis with validation loops, document processing with human sign-offs<\/span><\/li>\n<\/ul>\n<p><b>Use CrewAI when:<\/b><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Task decomposes naturally into specialist roles &#8211; researcher, writer, editor, analyst, validator<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Non-engineers need to read or contribute to agent architecture definitions<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Speed to prototype matters; your team will be judged on how quickly you can demonstrate value<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Use cases: content automation, research synthesis, operations automation, marketing workflows<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">You can absorb token overhead, or workflow volumes are moderate enough that cost efficiency is secondary to velocity<\/span><\/li>\n<\/ul>\n<p><b>Use AutoGen when:<\/b><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Agents need to write and execute code as a core workflow &#8211; data analysis, automated code review, and report generation<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Human participants need to be in the agent loop dynamically<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Operating in a Microsoft-heavy environment (Azure AI, M365, Teams)<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Complex multi-agent conversations where emergent collaboration adds value, a structured graph cannot capture<\/span><\/li>\n<\/ul>\n<p><b>Use LangChain Core when:<\/b><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">RAG pipelines or integration-heavy applications leveraging 1,000+ connectors<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Single-agent or simple chain-based workflows without complex state management requirements<\/span><\/li>\n<\/ul>\n<p><a href=\"https:\/\/multiqos.com\/contact-us\/\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19307\" src=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Get-a-production-ready-AI-agent-built-on-LangGraph-CrewAI-or-Microsoft-Agent-Framework-matched-to-your-stack-compliance-and-budget.webp\" alt=\"Get a production-ready AI agent built on Framework matched to your stack, compliance, and budget\" width=\"1400\" height=\"418\" srcset=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Get-a-production-ready-AI-agent-built-on-LangGraph-CrewAI-or-Microsoft-Agent-Framework-matched-to-your-stack-compliance-and-budget.webp 1400w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Get-a-production-ready-AI-agent-built-on-LangGraph-CrewAI-or-Microsoft-Agent-Framework-matched-to-your-stack-compliance-and-budget-430x128.webp 430w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Get-a-production-ready-AI-agent-built-on-LangGraph-CrewAI-or-Microsoft-Agent-Framework-matched-to-your-stack-compliance-and-budget-1024x306.webp 1024w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Get-a-production-ready-AI-agent-built-on-LangGraph-CrewAI-or-Microsoft-Agent-Framework-matched-to-your-stack-compliance-and-budget-150x45.webp 150w\" sizes=\"auto, (max-width: 1400px) 100vw, 1400px\" \/><\/a><\/p>\n<h2><b>Real Implementation Architectures: AI Agent Framework Use Cases<\/b><\/h2>\n<p><span style=\"font-weight: 400;\">Theory travels. These four architectures have been validated in production. Framework selection also depends on the <\/span><a href=\"https:\/\/multiqos.com\/blogs\/ai-agent-use-cases\/\"><span style=\"font-weight: 400;\">use case of the AI Agent<\/span><\/a><span style=\"font-weight: 400;\"> you are developing. Here are some of the use cases and best framework picks for those.<\/span><\/p>\n<p><span style=\"font-weight: 400;\"><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-19305\" src=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Real-Implementation-Architectures_-AI-Agent-Development-Framework-Use-Cases.webp\" alt=\"\" width=\"2048\" height=\"1386\" srcset=\"https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Real-Implementation-Architectures_-AI-Agent-Development-Framework-Use-Cases.webp 2048w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Real-Implementation-Architectures_-AI-Agent-Development-Framework-Use-Cases-430x291.webp 430w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Real-Implementation-Architectures_-AI-Agent-Development-Framework-Use-Cases-1024x693.webp 1024w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Real-Implementation-Architectures_-AI-Agent-Development-Framework-Use-Cases-1536x1040.webp 1536w, https:\/\/multiqos.com\/blogs\/wp-content\/uploads\/2026\/06\/Real-Implementation-Architectures_-AI-Agent-Development-Framework-Use-Cases-150x102.webp 150w\" sizes=\"auto, (max-width: 2048px) 100vw, 2048px\" \/><\/span><\/p>\n<h3><b>Use Case 1: Customer Service Agent (Best Pick: CrewAI)<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">Customer service is where CrewAI really shines. Why? Because CrewAI lets you assign each agent a clear job, like you would with a real team.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">A normal setup uses four agents working together:<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">One reads the customer&#8217;s message and figures out what they&#8217;re asking about.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">One digs through your help docs and old support tickets to find useful info.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">One writes the reply that the customer will actually see.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">One checks that reply before it goes out, making sure the tone is right and nothing breaks company policy.<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">The best part is that these roles are written in plain English. Your customer support team can read them and understand exactly what each agent does. No coding background needed.<\/span><\/p>\n<h3><b>Use Case 2: Data Analysis Agent (Best Pick: AutoGen)<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">If your agent needs to write code, run it, and read the output, AutoGen is the one you want. That&#8217;s its strong suit.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Here&#8217;s how it could work for a finance team. Someone types a question in plain English, like &#8220;Compare Q1 2026 revenue with Q1 2025, flag any region that dropped more than 15%, and make me a chart.&#8221; That request goes to the DataAnalystAgent.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">The agent writes a SQL query. Then it hands the query off to a second agent called the ExecutorAgent, which actually runs it against the data warehouse. The results come back. The DataAnalystAgent reads them, writes some Python code to build the chart, and passes that to the ExecutorAgent again.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">The chart gets made. The Data Analyst Agent looks at it, writes a short summary, and that&#8217;s the output. This write, run, check cycle keeps going on its own. It only stops when the result is good enough (based on rules you set), or when the agent runs into something confusing and pings a human for help.<\/span><\/p>\n<h3><b>Use Case 3: Workflow Automation Agent (Best Pick: LangGraph)<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">Some workflows are complex. They touch a bunch of different systems. They need a human to approve things at certain points. They have to keep working even if a server crashes. And every step needs to be logged for compliance. That&#8217;s exactly the kind of problem LangGraph was built for.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">The entire process is like this:\u00a0<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Pull in the document<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Find the important clauses<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Score the risk<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Check it against company policy<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Ping the legal team if the risk is too high<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Wait for someone to approve it<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Send it to whoever needs to sign<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Handle whatever comes after approval.<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">In LangGraph, each step of a process is a node. Each &#8220;if this, then that&#8221; moment is a conditional edge. And the whole thing carries a state, basically a notebook that tracks the document, the clauses pulled out, the risk score, who approved what, and the audit log.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Suppose the system crashes halfway through a review. Postgres has already saved a checkpoint. When everything comes back up, the workflow automation picks up right where it left off. The lawyer who was reviewing the contract doesn&#8217;t lose any work. And the audit trail stays intact, no matter what happened with the servers.<\/span><\/p>\n<h3><b>Use Case 4: Custom Pipeline Agent (Best Pick: LangChain Core)<\/b><\/h3>\n<p><span style=\"font-weight: 400;\">Sometimes you don&#8217;t want a framework making decisions for you. You want the parts. LangChain Core is the toolbox underneath everything else, so when your problem doesn&#8217;t fit a neat &#8220;team of agents&#8221; or &#8220;graph of nodes&#8221; shape, this is what you reach for.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Picture a content team that runs the same job over and over. A blog draft comes in. It needs to be summarized, checked against a brand voice guide, tagged with the right keywords, and rewritten if it misses the mark. None of that needs four agents debating each other. It needs one clean chain that does the steps in order, every time, the same way.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Here&#8217;s how it goes. The draft hits a prompt template that wraps it with instructions. That template feeds into a model. The model output runs through an output parser that pulls the structured bits you actually care about, like the tone score and the keyword list. Then a small piece of routing logic looks at that tone score. Above the threshold? Ship it. Below? Loop it back through a rewrite prompt and run it again.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">The whole thing is wired together with LCEL, the LangChain Expression Language. You chain the pieces with a pipe, like Unix commands: prompt | model | parser. Each piece is swappable. Don&#8217;t like the model? Swap it. Want a different parser? Drop it in. Nothing else breaks, because every component speaks the same interface.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">And that&#8217;s the real pull of LangChain Core. It doesn&#8217;t decide your architecture for you. You get the prompt templates, the model wrappers, the parsers, the retrievers, the memory objects, and you bolt them together however your problem demands. CrewAI gives you a team. LangGraph gives you a graph. LangChain Core gives you the raw building blocks, and trusts you to know what you&#8217;re building.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Worth noting: the other three frameworks all sit on top of LangChain Core anyway. So learning it isn&#8217;t a detour. It&#8217;s the foundation the rest stand on.<\/span><\/p>\n<h2><b>Conclusion<\/b><\/h2>\n<p><span style=\"font-weight: 400;\">There&#8217;s no winner here. The right framework is the one your team can actually run without losing sleep.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Think about it this way. A team using LangGraph but barely understanding it will mess up more than a team using CrewAI and knowing it inside out. Power doesn&#8217;t help if nobody on your team can drive the thing. Choosing among the AI agent frameworks is very subjective, depending on use cases, cost, and team knowledge.<\/span><\/p>\n<p><b>So how do you pick?<\/b><\/p>\n<p><span style=\"font-weight: 400;\">Start with what your business needs. If you work in a regulated space, like banking or healthcare, where every step has to be logged, and every approval has to be on record, LangGraph is the one. The learning curve is steep. You don&#8217;t get a choice. That&#8217;s just the cost of doing it right.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">But if your team has two weeks to ship something and the rules around it are loose, CrewAI gets you there fast. Speed is a real advantage when nobody is watching the audit trail.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">If you&#8217;re not sure which approach is right for your use case, working with an <\/span><a href=\"https:\/\/multiqos.com\/ai-agent-development-services\/\"><span style=\"font-weight: 400;\">AI agent development services provider<\/span><\/a><span style=\"font-weight: 400;\"> can help. A good provider will evaluate your business requirements, compliance needs, timeline, and scalability goals, then recommend the framework that delivers the best long-term value; whether that&#8217;s LangGraph for control and governance or CrewAI for rapid development and deployment.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Here&#8217;s the short version to remember:<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">LangGraph, when control and compliance come first.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">CrewAI when speed and team-style collaboration matter most.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Microsoft Agent Framework (the new home for AutoGen), when your stack is Microsoft, and your agents need to run code.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">LangChain Core is when you just need to connect a lot of tools together.<\/span><\/li>\n<\/ul>\n<p><span style=\"font-weight: 400;\">Pick the one your team can run pr handle easily and efficiently, not the one that looks best on a slide deck.<\/span><br \/>\n<script type=\"application\/ld+json\">\n{\n  \"@context\": \"https:\/\/schema.org\",\n  \"@type\": \"FAQPage\",\n  \"mainEntity\": [{\n    \"@type\": \"Question\",\n    \"name\": \"Can I switch frameworks after starting development?\",\n    \"acceptedAnswer\": {\n      \"@type\": \"Answer\",\n      \"text\": \"Yes, but it is expensive. Core abstractions, state management, agent definitions, and coordination patterns do not transfer cleanly. A mature LangGraph implementation cannot be migrated to CrewAI without significant rewriting. Front-load the evaluation with a structured decision matrix. The cost of a wrong choice compounds as your codebase grows.\"\n    }\n  },{\n    \"@type\": \"Question\",\n    \"name\": \"Does my team need ML expertise to use these frameworks?\",\n    \"acceptedAnswer\": {\n      \"@type\": \"Answer\",\n      \"text\": \"No. CrewAI is accessible to strong software engineers without ML backgrounds. LangGraph requires comfort with state machine design and async execution. Microsoft Agent Framework requires Docker and containerization comfort. None requires ML research expertise. What all four require: engineers comfortable with prompt engineering, debugging non-deterministic systems, and designing eval frameworks for LLM outputs.\"\n    }\n  },{\n    \"@type\": \"Question\",\n    \"name\": \"What replaced AutoGen?\",\n    \"acceptedAnswer\": {\n      \"@type\": \"Answer\",\n      \"text\": \"Microsoft Agent Framework. As of April 2026, both AutoGen and Semantic Kernel are in maintenance mode (bug fixes only, no new features). New code targets Microsoft. Agents. AI in .NET and Python.\"\n    }\n  },{\n    \"@type\": \"Question\",\n    \"name\": \"Do LangGraph, CrewAI, and Microsoft Agent Framework support MCP and A2A?\",\n    \"acceptedAnswer\": {\n      \"@type\": \"Answer\",\n      \"text\": \"Yes. MCP (Model Context Protocol) is the default tool-calling standard across the industry in 2026, with native support in Anthropic, OpenAI, Google, and Microsoft flagship models. A2A (Agent-to-Agent) is the cross-framework agent collaboration standard, governed by the Agentic AI Foundation.\"\n    }\n  }]\n}\n<\/script><\/p>\n","protected":false},"excerpt":{"rendered":"<p>60% of Fortune 500 companies are now running CrewAI in some form of production. That same year, LangGraph powers the majority of documented enterprise agent deployments requiring durable, stateful execution. Two frameworks. Wildly different adoption stories. Both are telling a version of the truth. The AI agent frameworks market has fragmented faster than most enterprises [&hellip;]<\/p>\n","protected":false},"author":4,"featured_media":19303,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[32],"tags":[],"class_list":["post-19298","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-ml"],"acf":[],"_links":{"self":[{"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/posts\/19298","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/comments?post=19298"}],"version-history":[{"count":8,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/posts\/19298\/revisions"}],"predecessor-version":[{"id":19312,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/posts\/19298\/revisions\/19312"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/media\/19303"}],"wp:attachment":[{"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/media?parent=19298"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/categories?post=19298"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/multiqos.com\/blogs\/wp-json\/wp\/v2\/tags?post=19298"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}