Personal Agent to Super-agent
Personal Agent to Super-agent: Beyond One Session
- Author
- DeepSeekAgent.io Editorial Team
- Published
- Updated
A Personal Agent answers “does this Agent continue to understand me?” It has a stable identity, preferences, memory, tools, and scheduled work. Super-agent asks another question: when messages and tasks arrive together, who prioritizes them, separates the work, and keeps several execution units isolated?
These are not competing product labels. They describe different layers: Personal Agent is about a durable relationship; Super-agent is about continuous coordination.
What a Personal Agent already solves
A mature Personal Agent can have editable preferences and memory, access to personal tools, scheduled wakeups, approval gates, and consistent behavior across Sessions. When work revolves around one person and one goal at a time, a dedicated DSH Profile and transparent workspace may be enough.
What breaks in one Session
Putting email triage, code changes, team replies, and research into one Session creates four problems:
- unrelated files and rules contaminate context;
- long work blocks new communication or gets interrupted by it;
- one high-privilege Session accumulates every tool and account;
- after compaction or restart, individual task state is hard to recover.
Opening more Sessions helps isolation but hands coordination back to the person, who must remember where every task lives.
Super-agent adds a scheduling layer
DeepSeek Harness places a Router between people and Workers. The Router receives Triggers from GitHub, Feishu, or other Sources, uses Policy to create Tickets, and dispatches each Ticket to an isolated Executor.
Personal memory still matters, but it no longer holds every task's state. Ticket owns work state, Worker owns execution context, Mandate owns authority, and Subscription sends later events to the right unit.
From assistant to a small team
| Stage | Main capability | Best fit |
|---|---|---|
| One-shot Agent | one conversation, one goal | temporary tool work |
| Personal Agent | memory, preferences, schedules | long-term service for one person |
| Multi-Session Agent | manually opened parallel Sessions | a few concurrent projects |
| Super-agent | Router, Policy, Tickets, Workers | continuous events and task dispatch |
Super-agent behaves more like a small virtual team: one front door receives requests, several Workers execute, and every task retains its own record and delivery channel.
Signals that it is time to upgrade
One Session is becoming insufficient when several issues or team threads need attention at once, new messages regularly interrupt long work, users spend time polling Agent status, tasks need different directories or identities, restarts lose the resumption point, or proactive detection is valuable but unrestricted action is not acceptable.
Without these pressures, a well-bounded Personal Agent is simpler to operate.
Where memory should live
Keep three types of state separate:
- user memory: preferences, identity, and long-term goals in the Personal Agent layer;
- task state: inputs, progress, questions, and delivery in the Ticket;
- execution state: directories, branches, commands, and model turns in Assignment and Worker.
Do not move everything into the Router prompt. The Router needs dispatch summaries and policy, Workers need their Ticket inputs, and durable preferences should remain editable data.
Initiative must become policy
Personal-agent requirements often say “be more proactive.” A Super-agent needs executable rules. The prototype offers off, observe, and auto acceptance modes so a team can inspect judgments before enabling automatic work.
A useful proactive rule names the Source and scope, qualifying events, people who can authorize work, allowed changes, escalation points, and the Channel for results.
A practical evolution path
- stabilize Personal Agent memory and approvals;
- split distinct responsibilities into separate DSH Sessions;
- select one continuous Source, such as one GitHub repository;
- validate Router decisions in
observe; - automate one repetitive, low-risk rule;
- add team chat, more repositories, and cross-source work last.
Super-agent should reduce the attention spent moving among Sessions, not require a person to surrender more control.
Related reading
- DeepSeek Harness Personal Agent tutorial: Profile to scheduled tasks
- DeepSeek Personal Agents: identity, memory, automation, and permissions
- DeepSeek Super-agent architecture: Router, Workers, and continuity
- dsh-superagent security: accounts, authority, and execution isolation