LobeHub: A Multi-Agent Platform That Plans, Codes, Tests, and Reviews Its Own Work
It turns Super-Agent theory into an auditable, productized workflow: agents produce run logs, screenshots, and test records, and a human only approves or bounces the final result. For teams weighing agent orchestration, LobeHub is a working reference that already ships its own product on itself.
LobeHub, the 80k-star open-source project formerly known as LobeChat, has become a multi-agent collaboration platform that puts the abstract Super-Agent architecture into a working product. A developer tested it by building a 3D web game, "180 Seconds Before Power Outage." A supervisor agent (June, on Opus 5.5) split the task, assigned a Codex engineer to write the game and an acceptance engineer to verify it. All context, progress, and comments stayed attached to a Linear-style Task card, so the whole execution history was auditable.
The platform implements the architecture's key pieces: Router/Worker separation, event-driven triggering via IM bots with a two-tier wake policy (rules first, then model judgment), and six execution environments ranging from local device to cloud sandbox. A cloud agent can call a local daemon through a device gateway, keeping code and toolchain on a Mac mini while the agent loop runs in the cloud. LobeHub's CAO mechanism treats tasks as cards, letting one person manage a dozen agents through dashboards instead of chat.
The test also introduced Acceptance, LobeHub's newly released open-source tool for agent-side acceptance testing. It evaluates deliverables with real run results, screenshots, and logs; acceptance failure triggers rework rather than marking the delivery failed. The strongest proof of the system: the LobeHub team uses it internally, with a supervisor agent named Nasa dispatching coding agent Dario to fix a real issue and submit PR #21852.
LobeHub's strongest claim is dogfooding: the team develops LobeHub with LobeHub, so every feature you test is already load-bearing in their daily workflow. That's more convincing than any demo.
The two-tier wake policy makes the abstract Fast Path/Slow Path routing concrete: rules cut noise first, then a model decides whether to interject. It's configurable and auditable, which is why it works outside a whiteboard.
Acceptance changes what 'done' means for an agent: it must produce run artifacts, screenshots, and logs, and failure is a normal loop back to development. Humans only veto at the end, which is a realistic division of labor.