The Stack Archipelago

Vercel-initiated open standard (1.0.0, August 2026) for packaging Agent Skills and MCP servers into distributable plugins. A plugin.json manifest + skills/ + mcp.json gives clients a common discovery and loading format, with steering committee members from AWS, Anysphere, GitHub, Microsoft, OpenAI, and Vercel. VS Code's Copilot August 2026 release added plugin installation following the 1.0 standard.

Microsoft's framework for building AI agents and applications, split into three parts: Studio for no-code prototyping, AgentChat for Python-based conversational agents, and Core for scalable multi-agent AI systems. Focused on cooperative task solving across heterogeneous agent personas.

Browser automation platform for agents offering three products: a Hosted Agents API for natural-language browser tasks, a managed Browser Infrastructure layer with SDK control, and an open-source Python library for building browser agents locally. Send a task, get a result from a managed browser.

Enterprise agent build-and-runtime platform where business and technical teams assemble AI agents and platform teams centrally govern them. CrewAI positions itself as the runtime for production agent backlogs, with role-based multi-agent crews that delegate tasks across specialists and shared observability for the rollout.

Open-source enterprise AI agent cloud — secure sandboxed environments with real-world tools (filesystem, shell, browser, networking) for running untrusted agent-generated code at scale. Marketed as the runtime layer that sits between LLM agents and the rest of the stack.

Open source AI orchestration framework from deepset (Berlin, since 2019) for production-grade agents. Composes retrieval, reasoning, memory, and tool-use pipelines against any LLM provider, with first-class integrations for OpenAI, Anthropic, Mistral, Hugging Face, Weaviate, Pinecone, and Elasticsearch. Used to ship enterprise RAG and agent systems at scale.

The original agent orchestration framework — chains, tools and LangGraph for stateful agents. Sibling product LangSmith reached GA on AWS in August 2026 with BYOC (Bring Your Own Cloud), letting enterprises keep observability data inside their own VPC.

AI agent framework offering document OCR and workflow automation. Its flagship LlamaParse product uses VLM-powered agents to parse complex documents (PDFs, tables, charts, handwriting) into structured, LLM-ready outputs for enterprise AI pipelines. Also ships indexing, retrieval, and agentic RAG primitives that connect private data to LLM apps.

Model Context Protocol — Anthropic's open standard wiring tools and data into agents.

OpenAI's lightweight, open-source (MIT) framework for multi-agent workflows. Provides agents with instructions, tools, guardrails, and handoffs; sandboxed long-running agents, realtime voice pipelines, agents-as-tools, human-in-the-loop, and built-in tracing. Provider-agnostic beyond OpenAI — supports the Responses and Chat Completions APIs plus 100+ other LLMs.

Open-source, framework-agnostic TypeScript toolkit for building AI applications and agents (now at ai-sdk.dev). Unified APIs cover text, image, speech, and video generation with streaming, provider fallbacks, and tool calling — letting teams swap model vendors without rewriting app code.