What is devswarm?

DevSwarm takes a single text prompt and produces a working full-stack application. You type something like "build me a habit-tracker app" and a team of AI agents built on open-weight models (GLM-5.2, Qwen3-Coder-480B, Kimi K2.x, and DeepSeek-V4) goes to work. One agent plans the project, two agents build the frontend and backend in parallel, and a critic agent reviews everything before it ships. The whole process takes three to six minutes using real model calls. The distinctive part of DevSwarm is how it watches itself. Every step each agent takes is recorded as a trace and sent to SigNoz, an open-source observability platform. The project ships with three dashboards and two alert rules that run on that trace data. If something goes wrong, like a model failing too often or the critic catching too many errors, a SigNoz alert triggers a webhook that wakes up a meta-agent called the Swarm Doctor. The Doctor reads the recent traces, diagnoses the problem in plain English, and adjusts which models handle which tasks, all on its own. Each generated application also ships already instrumented with its own OpenTelemetry setup, so the app's own traffic starts appearing in SigNoz seconds after it is created. A full generation produces traces, metrics like token counts and latency, and structured logs. When something is slow or broken, the answer is one trace away. The project includes a landing page with live stat cards and a Mission Control view showing a live swarm graph, span waterfalls with links into SigNoz traces, and token vitals. The repo also contains the swarm's contracts, a UI with no build step, and an observability folder with importable dashboards and alert rules. It was built for the Agents of SigNoz hackathon by WeMakeDevs, running July 20 to 26, 2026, in the AI and Agent Observability track.