clawhub/microsoft/agent-framework

• **Archetype**: Multi-agent framework with graph-based orchestration
• **Stars**: 9,365 (GitHub) / Discovered via ClawHub skill author jgarrison929 (fork)
• **Last pushed**: 2026-04-12
• **License**: MIT
• **Discovered**: 2026-04-12
• **Source**: clawhub-skills (indirect -- jgarrison929/security-auditor skill led to fork discovery)
• **Skills found**: 0 (framework repo)

Summary

Microsoft Agent Framework (formerly AutoGen/Semantic Kernel agents) is a comprehensive multi-language framework (Python + .NET) for building, orchestrating, and deploying AI agents. Key features:

• **Graph-based workflows**: Connect agents and deterministic functions using data flows with streaming, checkpointing, human-in-the-loop, and time-travel capabilities
• **Declarative agents**: YAML/JSON-based agent definition (declarative package)
• **Multiple orchestration patterns**: Sequential, parallel, group chat, handoff, nested
• **DurableTask integration**: Durable workflow execution with checkpointing
• **DevUI**: Interactive developer UI for agent development and debugging
• **A2A protocol support**: Agent-to-Agent communication standard
• **AG-UI integration**: Agent-UI protocol for frontend rendering
• **Lab package**: Experimental features including benchmarking and reinforcement learning
• **Multi-provider**: OpenAI, Anthropic (Claude), Azure, Bedrock, Ollama support

Architecture packages: core, orchestrations, declarative, durabletask, devui, a2a, ag-ui, lab, plus provider-specific adapters.

Assessment

**MEDIUM VALUE** -- While massive in scope, the framework is mostly infrastructure/SDK code rather than extractable process methodology. However, several architectural patterns are relevant to babysitter:

1. **Declarative agent definitions**: Microsoft's YAML-based agent definition pattern could inform babysitter's process definition format 2. **DurableTask integration**: Their approach to durable workflow execution with checkpointing is philosophically similar to babysitter's event-sourced replay 3. **DevUI**: Their interactive debugging UI is similar in concept to babysitter's observer dashboard 4. **Graph-based orchestration**: Their DAG-based workflow model could inspire enhancements to babysitter's parallel.all()/parallel.map() 5. **A2A protocol**: Agent-to-Agent communication standard could be useful for babysitter's multi-harness orchestration

The codebase is too large (1500+ files) for deep extraction but worth studying specific patterns.

Extraction Priority

**P2 -- Study for patterns, low urgency**

Processes

1. **Graph-Based Orchestration Methodology** (methodologies/): A methodology for defining agent workflows as DAGs with typed edges, checkpointing, and human-in-the-loop gates. Study their orchestrations package for patterns.

Plugin Ideas

1. **A2A Protocol Plugin**: Implement Agent-to-Agent protocol support in babysitter, enabling cross-framework agent communication (e.g., a babysitter-orchestrated agent talking to a Microsoft Agent Framework agent).

2. **Declarative Process Plugin**: YAML/JSON-based process definition format (inspired by Microsoft's declarative package) as an alternative to JavaScript process files.

Library Mapping

Plugin Marketplace Mapping