Live Production Demo:
- 🖥️ Interactive Web Dashboard: https://agent-trace-zeta.vercel.app/- ⚙️ FastAPI Swagger Docs: https://agenttrace-api-cdav.onrender.com/docs
An end-to-end observability SDK and dashboard for autonomous AI agent pipelines. It monitors multi-step tool calls, visualizes latency bottlenecks, and automatically repairs malformed LLM tool arguments at runtime without crashing workflows.
LLMs frequently hallucinate tool arguments during multi-step runs:
- Passing strings instead of floats (e.g.
"1200 INR"instead of1200.0) - Inventing key names (e.g.
"user_identifier"instead of"user_id") - Omitting required schema fields
Normally, these cause immediate runtime crashes. AgentTrace catches these failures and auto-repairs them at runtime.
- Decorator SDK: Python, Pydantic (Validates schema before tool run)
- Self-Healing Layer: Fast inference via Groq to repair payloads on failure
- Collector Backend: FastAPI with SQLite persistence (
traces.db) - Live Dashboard: Next.js, Tailwind CSS with Payload Diff Inspector
python -m uvicorn main:app --reload --port 8000
source & further reading
github.com — original article
Show HN: Badbox – A deterministic bad-pattern detector for codebases
Complete Production Webapp in Lean
Ogre Battle 64 Recompiled Project at 99.05%