dsh-maf

Microsoft Agent Framework (MAF) is the enterprise successor to Semantic Kernel, built around YAML-declarative agent and workflow definitions. A natural fit for letting an AI agent generate, validate, and execute those definitions, MAF previously had no adapter for DeepSeek Harness—so dsh agents could not read MAF YAML files out of the box, let alone run a full validate-then-execute loop inside a conversation. dsh-maf closes that gap by wrapping the MAF SDK into a set of dsh-native tools, so a single prompt can ask the agent to check the environment, validate a YAML file, run an agent or workflow, or scaffold a new declaration from scratch.

Five tools are exposed: maf_status inspects the MAF package version and Python environment; maf_validate parses a declarative-agent YAML through AgentFactory and reports schema issues; maf_run_agent loads the YAML, injects a prompt, and executes the agent via await agent.run(prompt); maf_run_workflow runs a workflow in the standard sample layout (workflow.yaml + main.py); and maf_scaffold generates a starter YAML skeleton for agents or workflows, saving the boilerplate work. Under the hood the engine calls AgentFactory.create_agent_from_yaml_path and defers to async execution, with a default timeout of 180 seconds to accommodate slower agent runs.

This plugin is well suited for developers and teams exploring MAF declarative orchestration who want a lower-friction path through dsh. It is especially handy when you are iteratively debugging YAML definitions, batch-generating agent templates, or need one-tap validation inside a dsh chat. Two caveats worth noting: running agents or workflows consumes your configured provider API quota, and the declarative-agents API is still marked experimental by Microsoft, so signatures may shift in upcoming releases.