AI agent harnesses, frameworks and runtimes

Low-level orchestration framework from LangChain for building, managing and deploying long-running, stateful agents as graphs.

Website

https://www.langchain.com/langgraph

Source

langchain-ai/langgraph · ★ 43,523 (as of 2026-10-20) — Python repo; LangGraph.js is a separate repo.

License

MIT

Languages

Python, JavaScript/TypeScript (LangGraph.js)

Hosting

Self-hosted library; LangSmith Deployment (Cloud, BYOC or self-hosted) for managed runtime

Best fit

Teams that want explicit, graph-level control over long-running stateful agents and may later want a managed deployment runtime.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Durable execution that resumes from where a run left off; human-in-the-loop via interrupts; short- and long-term memory; tracing and evals through LangSmith (separate product); MCP via LangChain's MCP adapter (langchain[mcp]); documented multi-agent patterns (subagents, handoffs, router).

Verified 2026-10-20 from: github.com · docs.langchain.com · docs.langchain.com · docs.langchain.com · api.github.com

Lightweight, provider-agnostic framework from OpenAI for multi-agent workflows built from agents, tools, handoffs, guardrails, sessions and tracing.

Website

https://openai.github.io/openai-agents-python/

Source

openai/openai-agents-python · ★ 30,144 (as of 2026-10-20) — Python repo; the JS/TS SDK (openai-agents-js) is separate.

License

MIT

Languages

Python, JavaScript/TypeScript

Hosting

Self-hosted library (runs in your process)

Best fit

Teams that want a small, opinionated agent loop with built-in tracing, and can add a durable-execution engine when runs get long.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Handoffs let an agent delegate to another agent; tool-approval pause/resume for human-in-the-loop; sessions (SQLAlchemy, SQLite, Redis, MongoDB, Dapr, encrypted); durable long-running runs via documented Dapr, Temporal, Restate and DBOS integrations; built-in tracing; MCP and hosted tools; sandbox, realtime and voice agents.

Verified 2026-10-20 from: github.com · openai.github.io · openai.github.io · openai.github.io · openai.github.io · api.github.com

Anthropic's SDK that exposes the Claude Code agent loop, built-in tools, permissions, sessions and hooks to Python and TypeScript applications.

Website

https://code.claude.com/docs/en/agent-sdk/overview

Source

anthropics/claude-agent-sdk-python · ★ 8,300 (as of 2026-10-20) — Python repo; the TypeScript SDK is a separate repo.

License

Use governed by Anthropic Commercial Terms of Service (GitHub reports MIT for the Python repo)

Languages

Python, TypeScript

Hosting

Self-hosted: the SDK supervises a Claude Code subprocess in your container/sandbox; Anthropic separately offers hosted Managed Agents

Best fit

Teams building Claude-based coding or computer-use style agents who want Claude Code's tools and permission model inside their own app.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Subagents; permission modes, rules and a canUseTool callback for approvals; hooks; sessions that can be resumed, forked and persisted to external storage (resume across hosts); file checkpointing/rewind; OpenTelemetry observability; external MCP servers and in-process SDK MCP servers for custom tools.

Verified 2026-10-20 from: github.com · code.claude.com · code.claude.com · code.claude.com · code.claude.com · code.claude.com · api.github.com

Google's open-source framework for building, debugging and deploying AI agents, from single agents to graph and multi-agent workflows.

Website

https://adk.dev/

Source

google/adk-python · ★ 21,913 (as of 2026-10-20) — Python repo; Java, Go and TypeScript ADKs are separate repos.

License

Apache-2.0

Languages

Python, TypeScript, Go, Java, Kotlin

Hosting

Self-host anywhere containers run; documented deployment to Agent Runtime (Google Cloud Agent Platform), Cloud Run and GKE

Best fit

Teams on Google Cloud or needing agents in several languages (incl. Java/Go) with a managed deployment path.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Modular multi-agent systems and graph workflows; human input and tool action confirmations; resumable agents (resume a stopped workflow); built-in logging, metrics and traces; MCP tools via McpToolset and agents exposed as MCP servers; A2A protocol.

Verified 2026-10-20 from: github.com · adk.dev · adk.dev · adk.dev · adk.dev · adk.dev · api.github.com

Microsoft's multi-language framework for production agents and multi-agent workflows; the stated successor to both AutoGen and Semantic Kernel.

Website

https://learn.microsoft.com/en-us/agent-framework/overview/

Source

microsoft/agent-framework · ★ 14,209 (as of 2026-10-20) — Go SDK lives in microsoft/agent-framework-go.

License

MIT

Languages

Python, C#/.NET, Go

Hosting

Self-hosted library; Foundry Hosted Agents for managed hosting on Microsoft Foundry

Best fit

.NET and Python shops, especially on Azure/Microsoft Foundry, and teams migrating from AutoGen or Semantic Kernel.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Graph-based workflows with sequential, concurrent, handoff and group-collaboration patterns; checkpointing, human-in-the-loop and time-travel; agent sessions and middleware; built-in OpenTelemetry tracing; MCP clients for tools and A2A interoperability; declarative YAML agents; DevUI.

Verified 2026-10-20 from: github.com · learn.microsoft.com · github.com · github.com · api.github.com

Python framework for multi-agent automation: Crews of role-based agents plus event-driven Flows, with CrewAI AMP as the commercial control plane.

Website

https://www.crewai.com

Source

crewAIInc/crewAI · ★ 59,879 (as of 2026-10-20)

License

MIT

Languages

Python

Hosting

Self-hosted open-source framework; CrewAI AMP Suite for managed deployment, observability and governance

Best fit

Teams prototyping role-based multi-agent automations in Python who may want a vendor-run control plane later.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Crews (autonomous agent teams) and Flows (event-driven workflows with managed state); checkpointing and async execution; human input / human-in-the-loop workflows; tracing and observability; MCP servers as tools and A2A support.

Verified 2026-10-20 from: github.com · docs.crewai.com · docs.crewai.com · docs.crewai.com · api.github.com

Python agent framework from the Pydantic team focused on type-safe agents, with a capabilities system, durable execution integrations and OpenTelemetry instrumentation.

Website

https://pydantic.dev/docs/ai/overview/

Source

pydantic/pydantic-ai · ★ 20,733 (as of 2026-10-20)

License

MIT

Languages

Python

Hosting

Self-hosted library; runs inside Temporal, DBOS, Prefect and other engines for durability

Best fit

Python teams that want typed, testable agents and to pick their own durable-execution engine.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Multi-agent patterns (agent delegation, programmatic hand-off, graphs); durable execution via TemporalDurability, DBOS and Prefect integrations, plus step persistence to save/resume/fork runs; deferred tools for human-in-the-loop tool approval; OpenTelemetry-native instrumentation (Pydantic Logfire or any OTel backend); MCP as a core capability.

Verified 2026-10-20 from: github.com · pydantic.dev · pydantic.dev · api.github.com

TypeScript framework for AI applications and agents, with workflows, memory, evals, observability and a local Studio.

Website

https://mastra.ai

Source

mastra-ai/mastra · ★ 28,917 (as of 2026-10-20)

License

Apache-2.0 core; code in ee/ directories under the Mastra Enterprise License (GitHub reports NOASSERTION)

Languages

TypeScript (Node.js, Bun, Deno, Cloudflare)

Hosting

Self-host on any Node.js-compatible environment (Mastra server, web framework or cloud provider); Mastra platform for observability, Studio and serving

Best fit

TypeScript/JavaScript teams who want agents and workflows in the same stack as their web app.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Supervisor agents that coordinate subagents; workflows with snapshots, suspend/resume, time travel and human-in-the-loop; traces, metrics and evals; MCPClient to consume MCP servers and MCPServer to expose agents, tools and workflows.

Verified 2026-10-20 from: github.com · mastra.ai · mastra.ai · mastra.ai · mastra.ai · mastra.ai · mastra.ai · raw.githubusercontent.com · api.github.com

Provider-agnostic TypeScript toolkit from Vercel for AI applications and agents, with tool-loop agents, subagents and durable workflow agents.

Website

https://ai-sdk.dev

Source

vercel/ai · ★ 27,400 (as of 2026-10-20)

License

Apache-2.0 (per LICENSE file; GitHub reports NOASSERTION)

Languages

TypeScript (Node.js; React, Next.js, Svelte, Vue, Angular UI integrations)

Hosting

Self-hosted library

Best fit

TypeScript/Next.js teams adding agents to web products.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: ToolLoopAgent and subagents a parent can delegate to; WorkflowAgent (@ai-sdk/workflow) for durable, resumable agents with automatic state persistence; tool execution approval (toolApproval); OpenTelemetry telemetry via @ai-sdk/otel; MCP tools.

Verified 2026-10-20 from: github.com · ai-sdk.dev · ai-sdk.dev · ai-sdk.dev · ai-sdk.dev · ai-sdk.dev · ai-sdk.dev · raw.githubusercontent.com · api.github.com

Open-source SDK and ready-made agent harness for Python and TypeScript that runs in-process with no hosted control plane; Amazon Bedrock is the default model provider.

Website

https://strandsagents.com

Source

strands-agents/harness-sdk · ★ 8,866 (as of 2026-10-20) — Monorepo (formerly strands-agents/sdk-python).

License

Apache-2.0

Languages

Python, TypeScript

Hosting

Self-hosted (in-process); documented deployment to AWS Lambda, EKS, Fargate, EC2 or Amazon Bedrock AgentCore Runtime

Best fit

AWS-centric teams that want an in-process agent harness with a serverless deployment path.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Multi-agent handoffs, swarms, graphs and workflows; sessions that persist and resume across restarts plus long-term memory; human approval gates for tool calls (interventions); tracing and evals; MCP tools; model-agnostic (Bedrock, Anthropic, OpenAI, Gemini, Ollama and more).

Verified 2026-10-20 from: github.com · strandsagents.com · api.github.com

Python framework and runtime (AgentOS) for running agents, teams and workflows as a service you host, with a web control plane.

Website

https://www.agno.com

Source

agno-agi/agno · ★ 42,814 (as of 2026-10-20)

License

Apache-2.0

Languages

Python

Hosting

Self-hosted AgentOS runtime on any container platform (Docker, Railway, AWS, GCP); data stored in your own database

Best fit

Teams that want to run a self-hosted multi-tenant agent platform with an API and control plane.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Teams of agents (incl. nested teams) and workflows; sessions, memory and traces stored in your database; human-in-the-loop (user confirmation, user input, external execution, admin approval); OpenTelemetry tracing and audit logs; exposes agents via a REST API and MCP server; MCP context providers. By default sends a telemetry event per agent run (vendor states prompts, messages and outputs are never sent); AGNO_TELEMETRY=false disables it.

Verified 2026-10-20 from: github.com · docs.agno.com · docs.agno.com · docs.agno.com · docs.agno.com · api.github.com

Hugging Face's minimal Python library for agents that write and execute their actions as code, with tool-calling agents also supported.

Website

https://huggingface.co/docs/smolagents/index

Source

huggingface/smolagents · ★ 29,929 (as of 2026-10-20)

License

Apache-2.0

Languages

Python

Hosting

Self-hosted library; optional remote/sandboxed code execution

Best fit

Developers and researchers who want a small, readable code-acting agent loop.

• Multi-agent: yes
• Durable state / resume: unknown
• Human-in-the-loop: unknown
• Tracing / observability: yes
• MCP support: yes

Details: CodeAgent and ToolCallingAgent; multi-agent hierarchies (managed agents); tools from any MCP server, LangChain or the Hugging Face Hub; share agents to the Hub; run inspection with OpenTelemetry; model-agnostic. Durable state and human-in-the-loop not verified.

Verified 2026-10-20 from: github.com · huggingface.co · huggingface.co · api.github.com

Stateful agent harness (formerly MemGPT) for agents with persistent memory, identity and skills that learn over time; usable from a CLI, the web, messaging channels or the Letta Agent SDK.

Website

https://www.letta.com

Source

letta-ai/letta-code · ★ 3,657 (as of 2026-10-20) — Current source repo. The original letta-ai/letta repo (25,193 stars) now points here and keeps the retired V1 server on an archive branch.

License

Apache-2.0

Languages

CLI via npm/uv/pipx; Letta Agent SDK for TypeScript

Hosting

Letta Cloud (default) or local / self-hosted App Server

Best fit

Teams building long-lived personal or shared agents where persistent memory is the core requirement.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: unknown
• MCP support: yes

Details: Persistent stateful agents with git-versioned memory (MemFS), memory blocks and skill learning; subagents and agents calling other agents; permission modes with interactive tool approvals; schedules/crons; hosted Letta MCP server (Letta Cloud) lets MCP clients create and message Letta agents. Tracing/observability not verified.

Verified 2026-10-20 from: github.com · github.com · docs.letta.com · api.github.com

Durable execution platform used as the orchestration layer under AI applications and agents; not an agent framework itself, but integrates with several.

Website

https://temporal.io/solutions/ai

Source

temporalio/temporal · ★ 23,833 (as of 2026-10-20) — Stars are for the Temporal server.

License

MIT

Languages

Go, Java, Python, TypeScript, .NET, Ruby, PHP, Rust (official SDKs)

Hosting

Self-hosted open-source server or Temporal Cloud (managed)

Best fit

Platform teams that need agent runs to survive crashes, deploys and multi-day waits, alongside an agent framework.

• Multi-agent: unknown
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: unknown

Details: Durable Execution: workflows hold state over long periods and recover from failures; human-in-the-loop (e.g. approving agent decisions); OpenTelemetry tracing of workflows and activities (e.g. Python SDK OpenTelemetryPlugin); listed integrations with OpenAI Agents SDK, Google ADK, Mastra and others; AI Cookbook. Multi-agent primitives and MCP support not verified.

Verified 2026-10-20 from: github.com · temporal.io · docs.temporal.io · docs.temporal.io · docs.temporal.io · api.github.com

TypeScript library for multi-agent networks with deterministic routing and MCP tooling, which gains retries, concurrency control and traces when run on Inngest.

Website

https://agentkit.inngest.com

Source

inngest/agent-kit · ★ 949 (as of 2026-10-20) — Last push to the repo per GitHub API: 2026-04-29.

License

Apache-2.0

Languages

TypeScript

Hosting

Self-hosted library served over HTTP; Inngest Platform for orchestration (Inngest Dev Server locally)

Best fit

TypeScript teams already on Inngest, or wanting event-driven, retryable agent networks.

• Multi-agent: yes
• Durable state / resume: yes
• Human-in-the-loop: yes
• Tracing / observability: yes
• MCP support: yes

Details: Agents combined into Networks with a Router and shared Network State; MCP servers as tools; human-in-the-loop via Inngest waitForEvent(); with Inngest: automatic retries of agent and tool calls (default 4 with exponential backoff), multitenancy/concurrency limits and step-by-step traces. Durability depends on deploying with Inngest.

Verified 2026-10-20 from: github.com · agentkit.inngest.com · agentkit.inngest.com · agentkit.inngest.com · agentkit.inngest.com · api.github.com