[Adapter] AutoGen / Semantic Kernel

[shreyas-lyzr]

Summary

Add adapters for Microsoft AutoGen and/or Semantic Kernel frameworks.

Details

Microsoft's AutoGen and Semantic Kernel are enterprise-grade agent frameworks. gitagent adapters would bridge the gap for teams using Microsoft's ecosystem.

Scope

Either or both:

• AutoGen: Multi-agent conversation framework
• Semantic Kernel: AI orchestration SDK (.NET, Python, Java)

Acceptance Criteria

• Export generates valid framework-specific code
• Run adapter executes agents correctly
• Documentation and examples added

Resources

• AutoGen docs
• Semantic Kernel docs

[krishvsoni]

Hi @shreyas-lyzr , I'd like to work on this

Approach: I'll start with the AutoGen adapter, following the existing adapter/runner pattern:

1. Adapter (src/adapters/autogen.ts): Generate Python code with AssistantAgent / GroupChat definitions:

   • SOUL.md + RULES.md → system_message
   • tools/*.yaml → Python function stubs with register_for_llm()
   • agents/ (sub-agents) → multiple ConversableAgent instances in a GroupChat
   • model.preferred → llm_config.config_list[].model
   • human_in_the_loop → UserProxyAgent with appropriate human_input_mode

2. Runner (src/runners/autogen.ts): Write temp .py file and spawn Python (same pattern as CrewAI runner)

3. Registration: Add case 'autogen' in export.ts, run.ts, and re-export from adapters/index.ts

4. Docs & Tests: Add docs/adapters/autogen.md with field mapping table + adapter tests

Will follow up with Semantic Kernel in a separate PR.

[shreyas-lyzr]

Apologies for the long silence. Still wanted, still open — but unclaimed. Reframing the ask in case it helps anyone considering taking it on:

AutoGen and Semantic Kernel are different enough that they should be separate adapters, not combined.

• AutoGen is conversation-driven (multi-agent group chat). Maps cleanly: agents/ → AutoGen Agent instances, skills/ → AutoGen tool functions, DUTIES.md → AutoGen role assignments. Likely 600–1000 lines of TS.
• Semantic Kernel is closer to OpenAI Agents SDK (skill-as-callable, planner-driven). Maps differently: skills/<name>/SKILL.md → SK plugin function, agent.yaml: model → SK kernel config. Different shape entirely.

If anyone wants to take one (not both), I'd prefer AutoGen first — bigger user base today, cleaner conceptual mapping. SK can follow.

Reference: src/adapters/crewai.ts is the closest existing adapter conceptually (multi-agent + role-based). Use it as a starting template.

good first issue is a stretch for this — help wanted is more honest, since either adapter is a multi-day chunk of work. If you'd like to take it on, comment here and I'll prioritize review.

Apologies again for the wait.

[shreyas-lyzr]

AutoGen and Semantic Kernel are both worth supporting given Microsoft's enterprise footprint. A few scoping notes:

AutoGen — the multi-agent conversation model maps well to gitagent's sub-agent concept. An adapter would need to handle: agent system prompts (SOUL.md + RULES.md → AssistantAgent system message), tool definitions (tools/*.yaml → registered function tools), and the ConversableAgent conversation termination conditions (maps loosely to max_turns).

Semantic Kernel — the plugin model is the natural mapping target. gitagent skills → SK plugins, tools → SK functions. SK supports .NET, Python, and Java — the adapter should target Python first (consistent with LangChain/LangGraph adapters) and note .NET as a follow-up.

Recommended approach — tackle AutoGen first. It is simpler (flat conversation model vs. SK's plugin/kernel architecture) and the multi-agent use case is a stronger selling point for gitagent. Open a separate issue for Semantic Kernel once AutoGen lands.

Note that PR #81 (LangChain) and PR #80 (LangGraph + DeepAgents) are in review now — once those merge, the adapter pattern will be well-established and this should be straightforward to implement.

[shreyas-lyzr]

With LangChain (#2) and LangGraph (#1) nearing resolution via PRs #81 and #80, AutoGen and Semantic Kernel are the logical next Microsoft ecosystem targets.

A few notes on scoping:

Autogen: The v0.4+ API (autogen-agentchat) uses a component model that maps reasonably well to gitagent — AssistantAgent for the main agent, tool registration via @tools, and sub-agent delegation via SelectorGroupChat or RoundRobinGroupChat. The gitagent skills would embed into the system_message.

Semantic Kernel: The Python SDK is more mature for this use case than .NET or Java for an export adapter. Agent plugins map to gitagent tools, KernelFunction maps to individual tool definitions. The multi-agent case would use AgentGroupChat.

Recommendation: start with AutoGen given it follows a similar pattern to the LangGraph adapter already in review. Semantic Kernel has more surface area and may warrant a separate issue.

Assigning @shreyas-lyzr to confirm priority and whether community PRs for these are welcome.