Adding some technical detail about our use of Huawei’s openJiuwen ecosystem:
We built Triviality’s WorkSwarm research harness around openJiuwen SwarmFlow. We chose it because its programmable Python workflows gave us explicit control over parallel exploration, agent handoffs, and review cycles. Separating orchestration from model execution also let us assign different models to individual roles without rewriting the research process. We packaged our protocol as a reusable Swarm Skill, including role definitions, coordination rules, and an executable workflow.
The coordinator develops three distinct approaches, and SwarmFlow’s parallel() API runs the researchers concurrently within each exploration round. Each researcher maintains an independent context before publishing findings to a shared discovery bank. The challenger returns structured feedback identifying questionable claims, mathematical evidence, unresolved gaps, and proposed tests. Reported refutations trigger abandonment of an approach, while repeated stagnation prompts a new direction. The shared record preserves failed attempts so replacement branches can learn from them.
Our Python runtime connects to a Node.js worker through structured JSON-lines events, with progress persisted in MongoDB for the dashboard. We embed the unchanged SwarmFlow engine at a pinned revision, check its integrity at startup, and use its journal to replay completed agent calls when inputs and workflow remain unchanged. Reviewed arguments advance to the proof writer and Lean verification, with compiler errors fed back into the research process. This gives us a concrete loop of exploration, criticism, revision, and checking, bounded by research limits while preserving partial findings.
Log in or sign up for Devpost to join the conversation.