Where Does Multi-Agent Work Actually Live?

A discussion post asks where shared project state, task ownership, and accountability should live when multiple AI agents work on the same project, noting Hugging Face already has work around agent frameworks, models, MCP, and agent runtimes. The post points to Sharkly.ai as taking the approach of placing agents and human teammates in the same project workspace, and asks builders using Hugging Face agents, smolagents, or open models where they currently keep shared state, tasks, and accountability.

As multi-agent systems become more capable, I’m wondering about a layer that seems easy to overlook: the workspace where the actual work is coordinated. An agent can reason, call tools, and complete a task, but once several agents are involved, we also need: Shared project context Clear task ownership Persistent state between tasks Parallel execution Traceable handoffs and decisions Human review when needed Hugging Face already has interesting work around agent frameworks, models, MCP, and agent runtimes. But should coordination of the actual project work live inside the agent framework, or should it be a separate layer? I’ve been looking at Sharkly.ai , which takes the latter approach by putting agents and human teammates into the same project workspace. For people building with Hugging Face agents, smolagents, or open models: where do you currently keep the shared state, tasks, and accountability when multiple agents work on the same project?