The Super-agent blueprint: DeepSeek's design and LobeHub's self-running agent team
Agent systems are leaving the single-session confine; persistent, event-driven teams with explicit delegation and acceptance are becoming the baseline. This setup is already running inside an 80k-star open-source project that uses its own agents to add capabilities for further agent collaboration.
DeepSeek's Super-agent design moves agents out of single-session chat into persistent, event-driven systems. Parallel workers each carry independent context and subscribe to message channels, so follow-up messages bypass the router. A watchdog watches for failures, and a database holds session state. The result is a system that accepts work at any time, keeps executing, and recovers from problems.
A working LobeHub setup demonstrates the pattern. One supervisor agent coordinates six execution agents with distinct roles and local tool access (Codex, Claude Code, and LobeHub's own agents). Delegation cards specify the target repo, baseline commit, allowed file changes, and acceptance criteria. A daily inspection task runs at 9am Beijing time, keeps per-round records in task comments, and has separate topics for each execution. Event triggers watch for keywords and let agents act without human prompting, and research tasks can be handed off through WeChat.
The practical lesson is that autonomy without boundaries turns agents into a nuisance. Every check needs explicit triggers, scope, and a clear point of return. The earlier multi-agent stack described in this context stalled because it lacked event-driven decision-making, so every task waited on a human. The fix is clearer handoffs, not smarter models. Using these systems well is closer to project management than prompt engineering; the 'team' is all AI, and the human still owns judgment and acceptance.
Autonomy in a Super-agent is not a default state; it requires explicit triggers, scope boundaries, and a clear point of return. Without those, proactive agents become noise.
The dogfooding case shows a positive feedback loop: the capability LobeHub's agent added was exactly what enables further delegation, so the tool compounds its own usefulness.
The author's distinction between repetitive execution and judgment suggests the value of agents lies in freeing humans for decisions about goals and acceptance, not in replacing them. This reframes agent skill as project management rather than prompt engineering.
DeepSeek's design here is consistent with the open-source philosophy presented at the Wuzhen Summit: open-sourcing both the Agent Harness and the inference framework, letting the community help validate them. For us engineering folks, this is a good thing — the foundation is already laid by others, and we can focus all our energy on business orchestration.