Skip to main content

Supported Frameworks

AgentVisor™ is designed as a framework-agnostic secure runtime for AI agents. Whether you're building conversational assistants, research agents, or complex multi-agent systems, AgentVisor provides the security isolation, policy enforcement, and durable execution your agents need.

Framework Support

FrameworkStatusNotes
LangGraph✅ SupportedFull integration with checkpointer, graph discovery, and streaming
CrewAI✅ SupportedMulti-agent crews with crewai.yaml discovery and MCP tool integration
Interactive✅ SupportedDirect command execution for agentvisor exec — shells, Claude Code, Aider
Google ADK✅ SupportedSession-scoped state with Thread State API; agent.yaml discovery; AgentVisorSessionService automatically wired
Strands Agents✅ SupportedAWS's code-first agent SDK; snapshot-based checkpointing; native MCP tool adapter; explicit framework.provider: "strands" selection (no discovery file)
Pydantic AI🗓️ RoadmapNative A2A + MCP; high production adoption; DBOS durability maps to AgentVisor checkpoint API
OpenAI Agents SDK🗓️ RoadmapLargest user base; native MCP; Sessions/Conversations API backed by AgentVisor store
Agno🗓️ RoadmapNative MCP + A2A + AG-UI; DB-backed sessions map to AgentVisor store API
LlamaIndex🗓️ RoadmapWorkflowCheckpointer maps to AgentVisor checkpoint API; strong RAG story
Smolagents🗓️ RoadmapHuggingFace ecosystem; code-agent paradigm well-suited to gVisor sandbox (strongest syscall isolation); native MCP
Microsoft Agent Framework🗓️ RoadmapSuccessor to AutoGen and Semantic Kernel; native MCP + A2A
Haystack🗓️ RoadmapPipeline-based model; YAML-defined pipelines for agent discovery

Framework Provider Architecture

AgentVisor uses a pluggable framework provider system. Each provider knows how to discover agent schemas and build the command to start the agent process inside the sandbox. The built-in providers are:

ProviderDiscovery FileDescription
langgraphlanggraph.jsonLangGraph Python agents with graph discovery and checkpointing
crewaicrewai.yamlCrewAI multi-agent crews with @CrewBase pattern
interactive(none)Direct command execution for agentvisor exec (shells, Claude Code, etc.) — selected via explicit framework.provider: interactive, or automatically as a fallback when agentvisor exec finds no other discovery file
adkagent.yamlGoogle ADK agents with session state backed by Thread State API
strands(none — explicit framework.provider: strands required)AWS Strands Agents with snapshot-based checkpointing and native MCP tool adapter

Because the sandbox is a Linux environment, any agent that can run on Linux can be supported by a framework provider. The provider handles schema discovery and command construction — the security, policy, and observability layers apply regardless of what runs inside.

See the Configuration Reference for framework configuration details.

Choosing a Framework

LangGraph is recommended when you need:

  • Stateful conversation threads with durable checkpointing via Temporal
  • Graph-based control flow with conditional edges and branching
  • Human-in-the-loop patterns with interrupt()/resume
  • LangGraph Agent Protocol API compatibility

CrewAI is recommended when you need:

  • Multiple specialized agents collaborating on a task
  • Role-based agent design with @CrewBase, @agent, @task decorators
  • Sequential or hierarchical task pipelines
  • Simple, declarative multi-agent workflows

Interactive is recommended for:

  • Running existing tools inside a policy-enforced sandbox (Claude Code, Aider, shells)
  • Wrapping CLI tools that aren't Python agents
  • Interactive terminal sessions via agentvisor exec

Google ADK is recommended when you need:

  • Gemini-native agents with ADK's built-in tool and session model
  • Session-scoped state automatically persisted across conversation turns
  • Fine-grained state scoping (session / user / app / temp) within one agent

Strands Agents is recommended when you need:

  • A code-first agent SDK (no YAML/JSON discovery file) with a broad choice of model providers
  • Native MCP tool support built into the framework itself
  • Multi-agent Graphs (DAG) or Swarms (handoff-based) alongside single agents
  • A confirmed production path to Bedrock via LiteLLM
  • Native human-in-the-loop interrupts

Roadmap

We're actively working to expand framework support. Future integrations will maintain the same security guarantees — sandbox isolation, policy enforcement, and credential brokering — while adapting to each framework's execution model.

Roadmap priorities are driven by three factors:

  • Protocol alignment — Frameworks with native A2A and MCP support (Google ADK, Pydantic AI, Agno, MS Agent Framework) unlock the full capabilities of AgentVisor's transport and gateway layers without adapter code.
  • Durability fit — Frameworks with pluggable session or checkpoint interfaces (LlamaIndex WorkflowCheckpointer, Agno storage backends) enable deep integration with AgentVisor's Temporal-backed checkpoint and store APIs.
  • Ecosystem reach — High-adoption frameworks (OpenAI Agents SDK, Pydantic AI) maximize the number of existing agents that can be secured with AgentVisor without rewriting.

Note on Microsoft frameworks: AutoGen is entering maintenance mode per Microsoft's announcement; all new development targets Microsoft Agent Framework (RC1: Feb 2026). Semantic Kernel's Python SDK is being unified into the same repo. AgentVisor roadmap tracks the successor rather than the predecessors.

If you're interested in a specific framework, contact us at support@manetu.com to discuss your use case.

Next Steps