AutoGen with PraisonAI

Use AutoGen v0.2 via the AutoGen family router or autogen_v2 direct alias in PraisonAI

PraisonAI integrates with the AutoGen family via a version router that picks the right adapter at runtime.

Which AutoGen Option to Choose

Only AutoGen v0.2 is working today. Both v0.4 and AG2 adapters are stubs that raise NotImplementedError.

Auto Version Selection

framework: autogen routes to a family adapter that picks the concrete version at runtime. Control it with the YAML autogen_version field (takes precedence) or AUTOGEN_VERSION:

framework: autogen
autogen_version: v0.2   # "auto" | "v0.2" | "v0.4"

# Force v0.2 (default behaviour)
export AUTOGEN_VERSION=v0.2
praisonai agents.yaml --framework autogen

# Explicit v0.2 alias — skip the router entirely
praisonai agents.yaml --framework autogen_v2

If no AutoGen variant is installed, the family router raises:

# No AutoGen variant installed → ImportError:
#   No AutoGen variant is available. Install with:
#     pip install 'praisonai-frameworks[autogen]' for v0.2
#     pip install 'praisonai-frameworks[autogen-v4]' for v0.4
#     pip install 'praisonai-frameworks[ag2]' for AG2

AutoGen v0.2

Original AutoGen v0.2 support — the only fully-working AutoGen adapter today.

1. Install

   pip install "praisonai-frameworks[autogen]"

2. Create YAML file

   framework: autogen
   topic: Create a movie script about cats on Mars
   
   roles:
     researcher:
       role: Research Analyst
       goal: Gather information about Mars and cat behavior
       backstory: Skilled researcher with focus on accurate information.
       tasks:
         research_task:
           description: Research Mars environment and cat behavior for {topic}
           expected_output: Research findings document with key facts

3. Run

   export OPENAI_API_KEY=sk-...
   praisonai agents.yaml --framework autogen

Pin v0.2 directly

Use autogen_v2 to skip the family router and get AutoGen v0.2 unconditionally:

framework: autogen_v2
topic: Create a movie script about cats on Mars

roles:
  researcher:
    role: Research Analyst
    goal: Gather information about Mars and cat behavior
    backstory: Skilled researcher with focus on accurate information.
    tasks:
      research_task:
        description: Research Mars environment and cat behavior for {topic}
        expected_output: Research findings document with key facts

praisonai agents.yaml --framework autogen_v2

AutoGen v0.2 Features

• Simple Architecture: UserProxyAgent + AssistantAgent with initiate_chats
• Configuration: Uses config_list format for LLM configuration
• Safe-by-default code execution: Opt-in via config.autogen.code_execution; Docker on by default — see AutoGen Code Execution
• Termination: Checks for "terminate" or "TERMINATE" in message content; human_input_mode now defaults to TERMINATE, overrideable via config.autogen.human_input_mode

Per-Agent LLM Override

Give each role its own model with llm:. It is layered over the shared config_list entry, so anything you don’t set (like api_key) is inherited.

framework: autogen
topic: AI agents

roles:
  researcher:
    role: Researcher
    goal: Find facts about {topic}
    backstory: Careful researcher
    llm: gpt-4o                    # string → {"model": "gpt-4o"}
    tasks:
      research:
        description: Research {topic}
        expected_output: Key facts
  writer:
    role: Writer
    goal: Write about {topic}
    backstory: Clear writer
    llm:                           # dict → overlays every key it sets
      model: llama3
      base_url: http://localhost:8000/v1
    tasks:
      write:
        description: Write a short article on {topic}
        expected_output: A 300-word article

AutoGen v0.4

AutoGenV4Adapter.implemented = False causes is_available() to return False, so framework: autogen and AUTOGEN_VERSION=auto will not route to it.

When implemented, AutoGen v0.4 will use autogen-agentchat + autogen-ext with OpenAIChatCompletionClient and RoundRobinGroupChat.

Track issue #1590 for implementation status.

AG2

AG2Adapter.implemented = False causes is_available() to return False, so AUTOGEN_VERSION=ag2 will log a warning and fall through.

Track issue #1590 for implementation status.

How the AutoGen Family Router Works

AutoGenFamilyAdapter (registered as autogen) is a router — it never runs directly. Instead it calls resolve() to get the concrete adapter, which the orchestrator then calls .run() on.

AutoGenFamilyAdapter.run() raises RuntimeError if called directly — this is intentional. Always go through the orchestrator.

Troubleshooting

Import Errors

When no AutoGen variant is installed and framework: autogen is used:

ImportError: No AutoGen variant is available. Install with:
  pip install 'praisonai-frameworks[autogen]' for v0.2
  pip install 'praisonai-frameworks[autogen-v4]' for v0.4
  pip install 'praisonai-frameworks[ag2]' for AG2

Fix: pip install "praisonai-frameworks[autogen]" for v0.2 (the only working option today).

NotImplementedError on v0.4 or AG2

NotImplementedError: AutoGen v0.4 adapter is not yet implemented. Use framework='autogen' (v0.2) or pin AUTOGEN_VERSION=v0.2.
NotImplementedError: AG2 adapter is not yet implemented. Use framework='autogen' (v0.2) or pin AUTOGEN_VERSION=v0.2.

Switch to framework: autogen or framework: autogen_v2. Both adapters are stubs pending issue #1590.

Pinning AutoGen versions

Setting AUTOGEN_VERSION=v0.4 or AUTOGEN_VERSION=ag2 (or YAML autogen_version: v0.4) now raises ImportError if the requested variant is not registered, instead of silently falling back to v0.2. Workflow-level autogen_version in config/YAML takes precedence over the env var. Use AUTOGEN_VERSION=auto (or omit it) to allow fallback.

Since autogen_v4 and ag2 adapters are unregistered stubs by default, --framework autogen resolves to v0.2 unless you install an entry-point plugin that registers autogen_v4 or ag2.

Unsupported field warning for function_calling_llm

framework='autogen_v2' ignores YAML field(s) ['function_calling_llm'] for agent 'Researcher'; these are only honoured by framework=praisonai.

AutoGen has no separate function-calling model. Remove function_calling_llm and set the role’s llm: instead.

cannot enforce safety field(s) error

ValueError: framework='autogen_v2' cannot enforce safety field(s) ['approval'] for agent 'Researcher'. Use framework: praisonai, or remove the field(s) to run ungated.

AutoGen adapters (autogen, autogen_v2, autogen_v4, ag2) only read llm, so an active approval: or guardrails: gate fails closed. Switch to framework: praisonai, or remove the field (or set approval: false) to run ungated. See Framework Safety Enforcement.

Unknown LLM provider prefix error

A provider/model with an unknown prefix and no base_url / OPENAI_API_BASE raises ValueError. Set an OpenAI-compatible endpoint or register the provider — see Provider Prefix Validation.

AUTOGEN_VERSION=v0.4 warns but continues

When AUTOGEN_VERSION=v0.4 is set but v0.4 is not installed (or is a stub), the router logs a warning and returns "autogen_v4" — the caller then hits NotImplementedError when .run() is invoked. This is expected behaviour; switch to v0.2.

Plugin authors: Framework adapters now accept dispatcher kwargs (tools_dict, agent_callback, task_callback, cli_config). See Framework Adapter Plugins for custom adapter development.

Best Practices

Related