Inspiration Multi-agent systems fail silently. When a swarm of agents breaks, you're left guessing which agent, which tool call, or which prompt caused it. Existing tools (LangSmith, Langfuse) proved this is a venture-scale need — we built the fast, visual, replayable version.
What it does telemetryZero is an observability + replay layer for multi-agent systems. It traces every agent, LLM call, and tool call onto a timeline with latency/token/cost, clusters failures (bad tool args, hallucinations, timeouts), scores each run with an LLM-as-judge, and lets you replay any run to watch the decision path unfold. It runs a live Researcher→Analyst→Writer→Judge swarm and traces it in real time.
How we built it Next.js 16 + React 19 + TypeScript + Tailwind on Vercel. A /api/run route orchestrates a multi-agent swarm, instrumenting each step into a structured trace. Live web search via Tavily; reasoning + eval via OpenAI. Custom timeline/replay engine — no charting library.
Challenges Designing a trace model that works for both seed and live runs, building a replay playhead that stays aligned across a variable-length timeline, and making the whole thing demo-proof with zero API dependency as a fallback.
Accomplishments A genuinely useful dev tool, shipped and deployed live in one day, with a real "wow" replay moment and live sponsor (Tavily) integration.
What's next Rollback snapshots, prompt-diff + fix-and-rerun, an aggregated agent DAG view, and OpenTelemetry ingest so any framework can pipe traces in.
Built With
- next.js
- openai
- react
- tailwindcss
- typescript
- vercel
Log in or sign up for Devpost to join the conversation.