AG2 Roadmap
Completed
- Agent
- Agent Harness (including knowledge, subtasks)
- LLM Providers
- OpenAI
- Anthropic
- Gemini (incl. Vertex AI)
- Ollama
- DashScope
- xAI
- Z.ai
- Function Tools
- Context Variables
- Dependency Injection
- System Prompt management
- Human-in-the-loop
- Middleware
- OpenTelemetry support
- Structured Output (static / callable / prompted / transformable)
- Built-in Tools
- Shell Tool
- AG-UI
- MCP Tools
- A2A
- A2UI
- Evaluation framework (offline mode)
- Evaluation (online mode + LLM-as-judge scorers + trajectory scorers + CLI)
- Multi-Agent Orchestration (aka Network)
- Multimodality inputs (images, audio, video)
- Multimodality streaming input: audio / video
- Multimodality output: audio / video output / images
- Live Agents
- STT & TTS
- Third-party contributions policy
- Local Code execution tool
- Messages aggregation
- Subagent delegation
- Dynamic subagents
- Background subagents
- First-class Usage tracking
- Skills support
- Agent as MCP server
In Progress
- Prometheus metrics
- Multi-Agent Network supports multimodality
- Multi-Agent Network supports streaming
- ACP Support (CLI agents orchestration)
- NLIP
Future Priorities
P1
- Scheduler support
- Repository development skills (code, structure, documentation, tests)
P2
- Deferred tools (Tool Search Tool)
- Builtin Memory toolkit
P3
- Builtin RAG toolkit
- Crawl4AI tool
- Allow models to configure tool search params
- MCP-UI
P4
- Use codegen to optimize tools configuration
- Msgspec serialization support
- TUI runtime
- Checkpoints and snapshots