AI agent frameworks compared

What an agent framework does

An AI agent framework is a library for building software in which a language model decides which tools to call, in what order, until a task is done. It supplies the agent loop, tool definitions, multi-agent patterns, state, and increasingly human approval and durable resume. It does not supply the business application: users, permissions, approval routing, the record of what was decided, or the pages people work in.

By late 2026 the leading frameworks have converged on a common feature set. Most support MCP tools, multiple model providers, multi-agent patterns, persistent state, and a way to pause for a person. The real differences are the core abstraction (graph, agent loop, roles, typed functions), the language, the cloud each aligns with, and what managed hosting its maker sells.

The frameworks at a glance

Graph and workflow frameworks

These frameworks make control flow explicit. You define steps and transitions, and the framework checkpoints state between them. They suit processes where the order of operations matters and must be inspectable.

LangGraph describes itself as a low-level orchestration framework and runtime for long-running, stateful agents. You define a graph over shared state; checkpointers save it per thread, and interrupt() pauses for a person. LangGraph 1.0 shipped on October 22, 2025, with a commitment to no breaking changes until 2.0. LangSmith adds tracing and evaluation, and LangSmith Deployment (formerly LangGraph Platform) hosts agents from LangGraph and other frameworks.

Microsoft Agent Framework is the direct successor to Semantic Kernel and AutoGen, built by the same teams. It combines agents, an opinionated Harness Agent for long multi-step tasks (planning, context compaction, file access, tool approval), and functional and graph-based workflows. Workflows checkpoint at the end of each superstep, with built-in storage in memory, on disk, or in Azure Cosmos DB. Version 1.0 shipped on April 3, 2026, for .NET and Python.

Google ADK (Agent Development Kit) is open source and available in five languages. It offers sequential, loop, and parallel workflow agents plus graph workflows, works with almost any model, and deploys to containers on your own infrastructure or to Agent Runtime on Gemini Enterprise Agent Platform, Cloud Run, and GKE. Resumable apps replay completed events after an interruption.

LlamaIndex offers prebuilt agents and Workflows, an event-driven, step-based model where each step consumes an event and emits the next. It suits teams already using LlamaIndex for retrieval over their documents.

Which agent frameworks support durable execution?

LangGraph, Microsoft Agent Framework, CrewAI, Mastra, and Google ADK persist state and resume. Pydantic AI integrates with durable engines directly, and the OpenAI Agents SDK has an official Temporal integration, generally available since March 23, 2026. Resuming a run is not the same as running a business action exactly once. Google ADK's docs say plainly that on resume, tools run at least once and may run more than once, so tools with side effects such as purchases must check for duplicates. LangGraph's docs give similar advice for code before an interrupt.

Agent-loop SDKs

These frameworks center on one capable agent loop and keep orchestration light. They suit teams that want minimal abstraction and plan to add structure in their own code.

OpenAI Agents SDK provides agents, handoffs (agents delegating to agents), guardrails, sessions, tracing, MCP tools, realtime voice agents, and sandbox agents that work in an isolated filesystem. It supports non-OpenAI models through adapters. It is OpenAI's recommended path for Agent Builder users: OpenAI notified developers on June 3, 2026 that Agent Builder and the Evals platform are deprecated, and both shut down on November 30, 2026.

Strands Agents is an Apache 2.0 open-source SDK from AWS for Python and TypeScript, described as a model-driven approach to building agents in a few lines of code. It supports Amazon Bedrock, Anthropic, OpenAI, Gemini, and other providers, with built-in MCP and multi-agent patterns (agents as tools, swarm, graph, workflow, A2A). It deploys to Bedrock AgentCore, Lambda, Fargate, EKS, and containers.

Pydantic AI brings typed outputs, dependencies, and tools to Python agents, with human approval of tool calls built in. Its durable execution story is distinctive: rather than its own runtime, it integrates with eight engines, including Temporal and DBOS as first-party integrations, so agents survive restarts on infrastructure you already operate.

smolagents from Hugging Face keeps its agent logic to about a thousand lines and favors CodeAgent, which writes its actions as code, executed in sandboxes such as Modal, E2B, or Docker. It suits research and code-heavy tasks.

Role-based multi-agent: CrewAI

CrewAI organizes work around roles. Each agent has a required role, goal, and backstory, and agents in a crew collaborate and delegate tasks. Flows wrap crews with state, events, and routing, and CrewAI's docs recommend starting production applications with a Flow. CrewAI AMP (Agent Management Platform) deploys each crew as a REST endpoint with tracing, a tool repository, and a no-code Crew Studio; a free Basic tier includes 50 workflow executions per month.

TypeScript-first: Mastra

Mastra is a TypeScript framework for agents and applications. Agents take instructions, a model string such as a provider and model name, and tools. Workflows can suspend() at a step and resume() from anywhere in your application, such as an HTTP endpoint, with snapshots stored in your configured storage provider so they survive restarts and deployments. LangGraph, the OpenAI Agents SDK, Google ADK, and Strands also offer TypeScript.

Optimizing programs: DSPy

DSPy, maintained by Stanford NLP under the MIT license, is less an agent framework than a way to program language models. You declare tasks as typed signatures, compose modules, and let optimizers tune prompts against a metric you define. It can sit alongside an agent framework to improve individual steps.

Which agent framework is best for enterprise?

Pyrana's view: choose by language, cloud, and the control model your engineers prefer, because capability gaps between the leaders are narrow and close quickly. Azure shops tend toward Microsoft Agent Framework, Google Cloud toward ADK, AWS toward Strands, and teams wanting explicit graphs toward LangGraph. Then plan for what none of them provides.

What every framework leaves you to build

Every framework above produces agents. None produces the enterprise application those agents work inside. That boundary is the subject of the pillar on agent framework vs runtime vs platform. In practice, a team shipping a governed application on any framework builds the following.

The OpenAI Agents SDK's own docs make one of these boundaries explicit: its run-state loaders do not authenticate the snapshot or the person submitting it, so the application must authorize the reviewer and prevent duplicate or replayed resumes. LangChain's docs note that self-hosted Agent Server has no default authentication. These are not shortcomings of the frameworks. They are a different layer.

Managed runtimes (LangSmith Deployment, CrewAI AMP, AWS Bedrock AgentCore, Agent Runtime on Google Cloud) close part of the gap: hosting, scaling, session state, and in some cases identity and policy for agents. The business application, its permission model, and its record of decisions remain yours.

Where Pyrana fits

Pyrana is not an agent framework. It is an agentic application platform: identity, authorization, storage, orchestration, approvals, audit, and context are built once, and each application is one App configuration plus a frontend experience. Agents are a mechanism inside the application, running on the Pyrana Harness with durable workflows on Temporal workers.

The rows in the table above are what Pyrana provides. Every call passes one fail-closed gate with an authorization check on the entity and an audit record written before dispatch. Tool approval policies freeze the exact invocation under a hash, route it to a named lane, check eligibility at decision time, and dispatch idempotently. Outputs are versioned, attributed resources. cortIQ holds company knowledge as classified Context Units with per-run citation accounting.

Frameworks and Pyrana connect through open interfaces. An agent built on any framework can work with a Pyrana App through the App MCP server, as the user, through the same gate. A framework agent exposed as an MCP server or REST endpoint can be a Pyrana tool source, and existing framework code can live in a client-owned extension. For head-to-head detail, see Pyrana vs LangGraph, Pyrana vs OpenAI Agents SDK, and Pyrana vs CrewAI.