Agent frameworks, multi-agent orchestration, and managed agent platforms.
OpenAI's current API for building agents — merges Chat Completions simplicity with Assistants-era tool use
OpenAI's recommended way to build agents going forward. Combines built-in tools (web search, file search, code interpreter, computer use) with a simpler, stateless-by-default request shape than the old Assistants API. This is where new integrations should start.
Managed agent platform with tools, files, and persistent threads — deprecated in favour of the Responses API
OpenAI has deprecated this in favour of the Responses API and set a sunset timeline. Built-in code interpreter, file search, and persistent threads were genuinely useful, but do not start new projects here — migrate existing integrations to the Responses API.
Anthropic's Model Context Protocol for universal tool integration
MCP is becoming the USB-C of AI tool integration. Open protocol lets Claude (and other models) connect to any data source or tool. Computer use for browser automation is unique.
Enterprise agent builder with Microsoft 365 and Azure integration
Best for Microsoft shops. Build agents that work across Teams, SharePoint, and Dynamics. Low-code builder for business users plus pro-code extensibility. Strong RBAC and compliance.
Microsoft's multi-agent conversation framework
Pioneered multi-agent conversation patterns. v0.4 rewrite (AgentChat) is cleaner and more composable. Good for complex workflows needing agent-to-agent collaboration. Strong research backing.
Role-based multi-agent framework with simple, intuitive API
Most approachable multi-agent framework. Define agents with roles, goals, and backstories — they collaborate on tasks. Great for rapid prototyping. CrewAI+ adds managed hosting.
Stateful agent workflows as controllable graphs from LangChain
Best for complex, stateful agent workflows with human-in-the-loop. Graph-based design gives explicit control over agent flow. checkpointing and replay built in. LangGraph Cloud for deployment.
OpenAI's production-ready successor to Swarm — agents, hand-offs, and guardrails
The production evolution of Swarm. Same lightweight hand-off pattern, but with built-in tracing, guardrails, and sessions that make it viable beyond prototypes. Works with the Responses API and, via LiteLLM-style adapters, other model providers. Good default for teams starting fresh on OpenAI-centric multi-agent work.
Lightweight experimental framework for agent hand-offs — superseded by the Agents SDK
The educational precursor to the OpenAI Agents SDK, which now covers everything Swarm did plus tracing and guardrails. Still fine for reading through minimal, readable hand-off code, but new projects should start on the Agents SDK instead.
Google Cloud's managed platform for building conversational agents
Best for GCP-native teams. Integrates with Google Search, datastores, and Gemini models. Dialogflow CX migration path. Good enterprise controls but less flexible than open frameworks.
AWS's managed agent service with knowledge bases and actions
Good for AWS shops needing managed agents with knowledge base retrieval. Multi-model support via Bedrock. Action groups enable tool use. Less flexible than open-source alternatives.
Python framework for building AI agents with memory and tools — rebranded and rebuilt from Phidata
The Phidata team rebranded to Agno with a full framework rewrite focused on runtime performance and multi-agent teams. Clean Python-first API, built-in memory and knowledge bases. Update old Phidata imports — the package name changed.
