---
title: "Agent frameworks & tools · the build layer, ranked · Agent Fieldbook"
url: https://agentfieldbook.org/software/frameworks/
description: "The build layer AI agents are assembled from: orchestration frameworks, execution tools and sandboxes, and SDK and observability harnesses, merged into one view ranked by GitHub stars. Each graded with its layer, kind, licence and place in the supply graph."
section: "Stack · Frameworks & Tools"
source: Agent Fieldbook — generated from the published page
---

# The build layer

**Frameworks · DAG / stateful · role-based multi-agent · conversational multi-agent · data / event-driven · chains / agents · plugin SDK · pipeline / RAG · typed agents · TS-native agents · Agent Development Kit · model-driven agents · multi-agent runtime · spec-driven workflow · Ontology-anchored orchestration + governance · Developer SDK · type-safe ontology access · LLM-authoring surface · Function-backed workflows · Tools · code sandbox (microVM) · MCP reference servers · browser automation · browser control · browser agent · dev sandbox · Harnesses · agent SDK · SDK (MCP-native) · TS AI SDK · Gemini SDK · Agents SDK (Workers) · model SDK · model gateway · observability / evals · observability / gateway · OTEL observability · observability / tracing · data/agent framework · agent SDK (code) · LM programming · stateful agents (memory) · app framework · agent SDK (Java) · autonomous platform · multi-agent (SOP) · agent framework · durable agents**

The concrete software an agent is assembled from: orchestration frameworks, execution tools + sandboxes, and SDK / observability harnesses. Filter by layer; ranked by GitHub stars (live). ◆ badges trace each into the graph.

## All software · ranked by GitHub stars Live data

## All software · 49

| Name | Layer | Kind | Stars | What it is | links· src |
| --- | --- | --- | --- | --- | --- |
| **langgraph** LangChain | Frameworks | DAG / stateful | 36,748 ★ | Build resilient agents. | [↗](https://github.com/langchain-ai/langgraph) |
| **crewAI** CrewAI | Frameworks | role-based multi-agent | 55,116 ★ | Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers | [↗](https://github.com/crewAIInc/crewAI) |
| **autogen** Microsoft | Frameworks | conversational multi-agent | 59,566 ★ | A programming framework for agentic AI | [↗](https://github.com/microsoft/autogen) |
| **llama_index** LlamaIndex | Frameworks | data / event-driven | 50,719 ★ | LlamaIndex is the leading document agent and OCR platform | [↗](https://github.com/run-llama/llama_index) |
| **langchain** LangChain | Frameworks | chains / agents | 141,250 ★ | The agent engineering platform. | [↗](https://github.com/langchain-ai/langchain) |
| **semantic-kernel** Microsoft | Frameworks | plugin SDK | 28,282 ★ | Integrate cutting-edge LLM technology quickly and easily into your apps | [↗](https://github.com/microsoft/semantic-kernel) |
| **haystack** deepset | Frameworks | pipeline / RAG | 25,847 ★ | Open-source AI orchestration framework for building context-engineered, production-ready LLM applications. Design modula | [↗](https://github.com/deepset-ai/haystack) |
| **pydantic-ai** Pydantic | Frameworks | typed agents | 18,266 ★ | AI Agent Framework, the Pydantic way | [↗](https://github.com/pydantic/pydantic-ai) |
| **mastra** Mastra | Frameworks | TS-native agents | 25,927 ★ | Mastra is the modern TypeScript framework for AI-powered applications and agents. | [↗](https://github.com/mastra-ai/mastra) |
| **adk-python** Google | Frameworks | Agent Development Kit | 20,514 ★ | An open-source, code-first Python toolkit for building, evaluating, and deploying sophisticated AI agents with flexibili | [↗](https://github.com/google/adk-python) |
| **harness-sdk** AWS | Frameworks | model-driven agents | 6,468 ★ | Build an agent harness and control it end-to-end. Open-source SDK for production AI agents in Python & TypeScript - any | [↗](https://github.com/strands-agents/harness-sdk) |
| **agno** Agno | Frameworks | multi-agent runtime | 41,044 ★ | Build, run, and manage agent platforms. | [↗](https://github.com/agno-agi/agno) |
| **GitHub Spec-Kit** GitHub | Frameworks | spec-driven workflow | - ★ | Open-source toolkit for spec-driven development with AI: a phase-per-command workflow (constitution → specify → plan → tasks → implement) that writes to a.specify/ folder tree | [↗](https://github.blog/ai-and-ml/generative-ai/spec-driven-development-with-ai-get-started-with-a-new-open-source-toolkit/) |
| **Palantir Ontology Agent Stack** Palantir Technologies | Frameworks | Ontology-anchored orchestration + governance |  | Enterprise agent stack anchored in Palantir's Ontology — a governed digital twin of enterprise data, business logic, and compliance rules. Three components: Agent Engine & SDK (concurrency, async, human approvals, fault tolerance in production workflows), Mindkit (fleet management + dynamic agent generation), Agent Studio (visual low-code agent builder). Agents traverse enterprise object links, execute predefined Action Types under the same governance a human user has, and complement LLM reasoning with deterministic simulation/optimisation tools. Exposes ontology primitives through Ontology MCP so external frameworks (ADK, Microsoft Agent Framework, Claude Desktop) can safely read and write the enterprise source of truth. | [↗](https://www.palantir.com/platforms/ontology/) |
| **Palantir Ontology SDK (OSDK)** Palantir Technologies | Frameworks | Developer SDK· type-safe ontology access |  | Multi-language toolkit giving developers direct, type-safe access to Foundry's Ontology from their own environment. Auto-generates types and functions for the developer's Ontology subset with IDE autocomplete; supports reading and writing objects, executing Action Types, subscribing to change streams (WebSocket / TypeScript OSDK), and running high-scale queries. Token-scoped access respects the same user permissions the Ontology enforces. Sits beneath the Agent Stack: OSDK is the SDK; Agent Engine, Mindkit and Agent Studio compose on top; Ontology MCP re-exposes the same primitives to external agents. | [↗](https://palantir.com/docs/foundry/ontology-sdk/overview/) |
| **Palantir AIP Logic** Palantir Technologies | Frameworks | LLM-authoring surface· Function-backed workflows |  | Palantir's authoring surface for LLM-powered Functions inside Foundry. Distinct from AIP Agent Studio (which composes multi-turn agents) — AIP Logic focuses on single-turn LLM invocations wrapped as Ontology Functions that Action Types, Agent Studio agents, and OSDK clients can call. Enforces the same governance, permissioning and ontology binding as any other Function-Backed Action, so an LLM step inside a workflow is not a governance escape hatch. Complements deterministic Functions authored in the same surface — the same file can host a regex extractor, an optimisation call, and an LLM reasoning step, all governed identically. | [↗](https://palantir.com/docs/foundry/aip/aip-logic/) |
| **E2B** E2B | Tools | code sandbox (microVM) | 12,887 ★ | Open-source, secure environment with real-world tools for enterprise-grade agents. | [↗](https://github.com/e2b-dev/E2B) |
| **servers** Anthropic | Tools | MCP reference servers | 88,181 ★ | Model Context Protocol Servers | [↗](https://github.com/modelcontextprotocol/servers) |
| **stagehand** Browserbase | Tools | browser automation | 23,416 ★ | The SDK For Browser Agents | [↗](https://github.com/browserbase/stagehand) |
| **playwright** Microsoft | Tools | browser control | 92,403 ★ | Playwright is a framework for Web Testing and Automation. It allows testing Chromium, Firefox and WebKit with a single A | [↗](https://github.com/microsoft/playwright) |
| **puppeteer** Google | Tools | browser control | 95,303 ★ | JavaScript API for Chrome and Firefox | [↗](https://github.com/puppeteer/puppeteer) |
| **browser-use** Browser Use | Tools | browser agent | 103,381 ★ | 🌐 Make websites accessible for AI agents. Automate tasks online with ease. | [↗](https://github.com/browser-use/browser-use) |
| **daytona** Daytona | Tools | dev sandbox | 72,266 ★ | Daytona is a Secure and Elastic Infrastructure for Running AI-Generated Code | [↗](https://github.com/daytonaio/daytona) |
| **openai-agents-python** OpenAI | Harnesses | agent SDK | 27,727 ★ | A lightweight, powerful framework for multi-agent workflows | [↗](https://github.com/openai/openai-agents-python) |
| **anthropic-sdk-python** Anthropic | Harnesses | SDK (MCP-native) | 3,717 ★ |  | [↗](https://github.com/anthropics/anthropic-sdk-python) |
| **ai** Vercel | Harnesses | TS AI SDK | 25,408 ★ | The AI Toolkit for TypeScript. From the creators of Next.js, the AI SDK is a free open-source library for building AI-po | [↗](https://github.com/vercel/ai) |
| **python-genai** Google | Harnesses | Gemini SDK | 3,835 ★ | Google Gen AI Python SDK provides an interface for developers to integrate Google's generative models into their Python | [↗](https://github.com/googleapis/python-genai) |
| **agents** Cloudflare | Harnesses | Agents SDK (Workers) | 5,235 ★ | Build and deploy AI Agents on Cloudflare | [↗](https://github.com/cloudflare/agents) |
| **openai-python** OpenAI | Harnesses | model SDK | 31,109 ★ | The official Python library for the OpenAI API | [↗](https://github.com/openai/openai-python) |
| **litellm** BerriAI | Harnesses | model gateway | 52,913 ★ | Python SDK, Proxy Server (AI Gateway) to call 100+ LLM APIs in OpenAI (or native) format, with cost tracking, guardrails | ◆ 1 [↗](https://github.com/BerriAI/litellm) |
| **phoenix** Arize | Harnesses | observability / evals | 10,449 ★ | AI Observability & Evaluation | [↗](https://github.com/Arize-ai/phoenix) |
| **helicone** Helicone | Harnesses | observability / gateway | 5,920 ★ | 🧊 Open source LLM observability platform. One line of code to monitor, evaluate, and experiment. YC W23 🍓 | ◆ 1 [↗](https://github.com/Helicone/helicone) |
| **openllmetry** Traceloop | Harnesses | OTEL observability | 7,280 ★ | Open-source observability for your GenAI or LLM application, based on OpenTelemetry | [↗](https://github.com/traceloop/openllmetry) |
| **langfuse** Langfuse | Harnesses | observability / tracing | 30,655 ★ | 🪢 Open source AI engineering platform: LLM evals, observability, metrics, prompt management, playground, datasets. Inte | ◆ 1 [↗](https://github.com/langfuse/langfuse) |
| **Claude Agent SDK** Anthropic | Harnesses | agent SDK | 7,600 ★ | Build autonomous agents with Claude Code capabilities: read code, edit files, run commands | [↗](https://github.com/anthropics/claude-agent-sdk-python) |
| **Google ADK** Google | Harnesses | agent SDK | 20,600 ★ | Code-first toolkit to build, evaluate and deploy agents; model-agnostic, tuned for Gemini/Vertex | [↗](https://github.com/google/adk-python) |
| **Microsoft Agent Framework** Microsoft | Harnesses | agent SDK | 12,100 ★ | Unified SDK and runtime for agents and multi-agent workflows; converges AutoGen + Semantic Kernel | [↗](https://github.com/microsoft/agent-framework) |
| **LlamaIndex** LlamaIndex, Inc. | Harnesses | data/agent framework | 50,800 ★ | Data framework for LLM apps over private data; document agents plus event-driven Workflows | [↗](https://github.com/run-llama/llama_index) |
| **Strands Agents** Amazon (AWS) | Harnesses | agent SDK | 6,500 ★ | Model-driven, model-agnostic SDK where the model reasons about which tools to call; OTel tracing | [↗](https://github.com/strands-agents/sdk-python) |
| **smolagents** Hugging Face | Harnesses | agent SDK (code) | 28,300 ★ | Barebones library whose agents think in code: the CodeAgent writes and runs Python, not tool JSON | [↗](https://github.com/huggingface/smolagents) |
| **DSPy** Stanford NLP | Harnesses | LM programming | 36,100 ★ | Programming, not prompting: declaratively compose and auto-optimise LM pipelines | [↗](https://github.com/stanfordnlp/dspy) |
| **Letta** Letta | Harnesses | stateful agents (memory) | 23,800 ★ | Platform for stateful agents with long-term memory that learn and self-improve across sessions | ◆ 4 [↗](https://github.com/letta-ai/letta) |
| **Genkit** Google | Harnesses | app framework | 6,200 ★ | Framework for AI-powered apps across JS, Go and Python; used in production at Google | [↗](https://github.com/genkit-ai/genkit) |
| **Spring AI** Broadcom (Spring) | Harnesses | agent SDK (Java) | 9,100 ★ | Application framework bringing Spring idioms to LLM apps and agents in the Java ecosystem | [↗](https://github.com/spring-projects/spring-ai) |
| **AutoGPT** Significant Gravitas | Harnesses | autonomous platform | 185,000 ★ | One of the original autonomous-agent projects; now a low-code platform for continuous agents | [↗](https://github.com/Significant-Gravitas/AutoGPT) |
| **MetaGPT** FoundationAgents | Harnesses | multi-agent (SOP) | 69,300 ★ | Multi-agent framework assigning SOP-based software-company roles (PM, architect, engineer) | [↗](https://github.com/FoundationAgents/MetaGPT) |
| **BeeAI Framework** IBM / Linux Foundation | Harnesses | agent framework | 3,300 ★ | Framework for production agents in Python and TypeScript, with tools, RAG and orchestration | [↗](https://github.com/i-am-bee/beeai-framework) |
| **Atomic Agents** Eigenwise | Harnesses | agent framework | 6,000 ★ | Schema-driven framework on Instructor + Pydantic composing agents as atomic, LEGO-like blocks | [↗](https://github.com/Eigenwise/atomic-agents) |
| **Julep** Julep AI | Harnesses | durable agents | 6,600 ★ | Framework for durable, composable agents: workflows that crash and resume, retry and record steps | [↗](https://github.com/julep-ai/julep) |

## Common questions

### What is the difference between an agent framework, a tool, and a harness?

A framework handles orchestration: the loop, the branching, the state carried between steps. A tool is something the agent acts through, like a code sandbox or a browser. A harness is the SDK that wraps the model call and records what happened. Most production stacks use one of each.

### Which AI agent framework should I choose?

Start from how much control flow you need. LangGraph suits agents with branching, retries and long-running state; lighter libraries like the OpenAI Agents SDK suit straightforward call-and-respond agents. Popularity is not fit, so pick for the shape of your problem. [(LangGraph)](https://langchain-ai.github.io/langgraph/)

### Why do agents need a code sandbox?

Because model-generated code can be wrong or hostile, and you do not want it running on your own machine. A sandbox gives it an isolated microVM or container with its own filesystem and restricted network. That boundary is what makes it safe to let an agent execute what it writes.

### Are AI agent frameworks open source?

Most of the widely used ones are, under MIT or Apache 2.0, including LangGraph, CrewAI, AutoGen and the OpenAI Agents SDK. A few carry custom terms. Check the LICENSE file before you commit, because the build layer is hard to swap out later.

### Do I need a framework to build an AI agent?

No. A loop that calls the model, checks for a tool call, runs it and feeds the result back is about fifty lines of code. A framework earns its place once you need persistence, retries, parallel branches, human approval steps or tracing across many runs.

## Next in the learning path

- [Models The models these frameworks call](https://agentfieldbook.org/software/models/)

- [MCP The tool-calling protocol](https://agentfieldbook.org/software/mcp/)

- [Agent tools The built-in capabilities these frameworks orchestrate](https://agentfieldbook.org/software/tools/)

- [Protocols The interop standards underneath](https://agentfieldbook.org/software/protocols/)

- [Examples Reference agents built on this layer](https://agentfieldbook.org/software/cookbooks/)

## Data

- [frameworks.json](https://agentfieldbook.org/data/frameworks.json)
- [tools.json](https://agentfieldbook.org/data/tools.json)
- [harnesses.json](https://agentfieldbook.org/data/harnesses.json)

Licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).