Who Moves First: Same Issue, Claude Didn't Reply, Super Agent Did
The dividing line between personal agents and super agents is who starts working first. Sean Chen tested this in the same GitHub repository: the DeepSeek open-source prototype needed three tags before it took the job, but in the reply test only it responded. The demo is not perfect, but the direction is worth watching.
Sean Chen posted a 19-minute video on November 15 about the difference between super agents and personal agents. In the video, he gives a very short criterion: who starts working first.
A personal agent is like an administrative assistant. It is always on, remembers you, can send and receive messages, and can be woken up by a scheduled task or a new email. Products like Hermes, Muse, and Grok Bot are all moving in this direction. The work is still initiated by a person, or by a cron job.
A super agent is more like a teammate. When a task is assigned, it may know before you do. DeepSeek open-sourced a prototype called dsh-superagent-prototype, and the usage is straightforward: tag it in a GitHub issue, the router opens a ticket, a worker does the work, and finally it returns to the original conversation and submits a pull request. What triggers it is an event, not a person.
The video uses a very small repository for the demo, called Tiny Invoices, a bill-counting tool. The format in the issue is wrong, and the amount displays incorrectly. Sean connects Waku Agent to the DeepSeek super agent, tags it in the issue, and asks it to fix the problem as soon as possible. The agent takes the task, changes the files, and opens a merge request. In the trajectory panel, you can see how the router assigns the task to the executor.
The full video is on YouTube, and the DeepSeek prototype is on GitHub.
The comparison experiment is in the same repository. One issue tags Claude, and another tags Waku Agent. After Claude is tagged, it directly starts a new task, reads the context, changes the code, and opens a pull request, without the router-and-ticket layer. The super agent takes another path: first to the router, the router decides whether to handle or ignore it, then opens a ticket and hands it to a worker.
The most interesting part is the comment-reply test. Under both issues, he left a single line: 'how did this go'. Claude did not reply. The super agent replied to both.
But the demo also has a very honest moment: one issue needed three tags before it was picked up. The video does not hide this. The prototype is not production software yet; triggers can fail, and speed still needs to improve. One commenter said that needing three tags before it takes the job shows both its potential and that it clearly needs someone watching it. Another commenter noticed that it took three tags to wake it up, and in the end only it replied to the comments.
Zooming out a bit, the form of agents has gone through several steps: agent harness, subagent, deep research, swarm, and then router and ticket. The earlier steps addressed how to organize agents so they can work in parallel, divide labor, research, and make decisions. Router and ticket move it one step forward: the agent stays always on, each task gets a ticket, and work starts as soon as an event arrives.
Why start with coding tasks? Sean's view in the video is that a lot of AI innovation is happening in coding right now; once coding tasks are done well, there is room to transfer that to design tasks and chief-of-staff-type tasks. Coding may be the foundation for more general agents.
Sean's own project, Waku Agent, takes another path: a local-first personal assistant. It breaks the agent into four pillars: Harness, Loop, Memory, and Eval/LLM-Ops. Memory is a SQLite file, the loop is about 95 lines of Python, a dashboard shows how each message flows through the harness, and eval includes both deterministic tests and LLM-as-judge. Waku Memory connects the same memory to Claude Code, Codex, Grok Bot, and Waku itself.

The Waku Agent repository is on GitHub, and shared memory is at waku.one.
This demo is not perfect. The prototype gets stuck, misses triggers, and needs three tags. But it puts a question on the table: we are moving from how one person uses a smart agent to how people and agents team up. Who starts working first determines whether the agent is a tool or a teammate.
发布时间: 2026-11-16 08:30