DeepSeek recently open-sourced an interesting project: dsh-superagent-prototype.
https://github.com/deepseek-ai/dsh-superagent-prototype
The idea is pretty straightforward: listen for GitHub events, have a Main Agent delegate tasks to Worker Agents, open PRs, and automatically follow up on CI failures and code reviews.
The Main Agent doesn't need to run continuously, and each delegated task has its own state and execution environment.
To me, this is a fairly conventional but practical approach. We've been building web backend services this way for years.
Everything is event-driven. No need for humans to constantly remind agents to keep working.
DeepSeek's Superagent essentially combines two independent loops to handle tasks and events over time:
- Runtime Loop: Handles events, task scheduling, and failure retries. Just regular software.
- Agent Loop: Handles the reasoning and tool execution within an individual agent's task.
In other words, it's basically an event-driven task management layer built on top of an Agent Harness.
For example, a Worker Agent opens a PR today. Tomorrow, someone leaves a review comment or CI fails. The system still knows which Worker owns that task and can route the event back to it for further action.
Nothing particularly magical here, but that's exactly why I think it's useful. The interesting part isn't making agents run forever. It's making their work persist across time and events.
Is DeepSeek reinventing the wheel a little? Probably. I suspect projects like GrokBot, Dots, and other personal agents are already exploring similar ideas.
But at least DeepSeek made the implementation open source so the rest of us can study it. That alone deserves some credit.
# AI #CodingAgent #MultiAgent