Agents Take Over an 84,000-Star Open-Source Company as LobeHub Opens Its Collaboration Platform

A hands-on report published by Chinese tech outlet QbitAI on November 9 describes how LobeHub, the team behind the open-source LobeChat project, has put AI agents inside its own engineering workflow, dispatching tasks, writing code and hunting bugs, and has released the source of its agent collaboration platform on GitHub, where the repository has drawn roughly 84,400 stars.

The report traces a concrete case from October 28. The team wanted agents to be able to specify which model to use when creating a task in its task tool and command-line interface, so a supervisor agent named Nasa opened task T-1149 and handed it to a coding agent called Dario. Dario implemented the feature, added tests and submitted pull request #21852; OpenAI's Codex review bot then flagged problems, Dario revised the code, and after several rounds a human reviewed and merged the change.

Handoffs also cross project boundaries. When the agent-foundry project needed custom neutral colour support in the lobe-ui component library, agents on both sides negotiated the requirement and the change directly, and with the release of lobe-ui 6.0.1 the upstream project picked up the new capability. One client data migration ran even longer: the agents worked for two days, three hours and 47 minutes, splitting the job into 60 tasks and completing 56 separate migrations.

The mechanism behind this is a deliberate separation of the agent loop, the session that stores task state, and the runtime that executes tools. If a user closes the chat, an agent can keep working in the background as long as the task lives in a persistent environment, and goals, progress, deliverables and acceptance records are kept in a shared workspace so the next agent knows what the previous one finished.

Human control is preserved at the end of the chain. LobeHub's Acceptance feature is multimodal: when an agent delivers work it attaches test results, screenshots and screen recordings, and a task cannot be marked complete until every check passes, with humans able to send it back for revision. In other words, the AI does the work, but sign-off stays with people.

The report also offers numbers. On 30 software-engineering tasks from FrontierHarness, the team's optimised agent harness using Kimi K3 completed 21, a 70 percent pass rate, at a median cost of 1.68 US dollars per successful task, which the article describes as ahead of Codex and Claude Code. A LobeHub developer says he has not started a local dev server in three months and estimates a three to five times efficiency gain; those figures are self-reported by the team and have yet to be reproduced at scale by outsiders.

On the timeline, LobeHub 2.0 in January 2026 introduced Agent Group and Memory, May brought CAO orchestration for chief agent operations, and the collaboration platform is now open-sourced alongside a Self-Driving Organizations framework graded L1 to L5. The team says it operates at L3 internally and is pushing some scenarios toward L4, where agents dispatch, hand off and accept work autonomously.

The approach echoes DeepSeek's recent exploration of a Super-Agent architecture, though LobeHub productised similar ideas earlier. QbitAI's own test, building an AI editorial team that produced six story ideas in under eight minutes and a news dashboard in about 30 minutes, offers a third-party scene check. The competitive question is shifting from how strong a single model is to who can string task decomposition, dispatch, review and acceptance into a reusable pipeline. What to watch next: whether the open-source community can reproduce these results, what the real costs are, and who is accountable when an agent gets it wrong.