Show HN: AgentTrace–Observability and runtime self-healing engine for AI agents

AgentTrace launched as an open-source observability SDK and dashboard that monitors multi-step tool calls in autonomous AI agent pipelines and automatically repairs malformed LLM tool arguments at runtime. The tool catches common LLM failures — passing strings instead of floats such as "1200 INR" instead of 1200.0, inventing key names like "user_identifier" instead of "user_id", and omitting required schema fields — and uses Groq-based fast inference to repair payloads before they crash workflows. The stack pairs a Python/Pydantic decorator SDK with a FastAPI and SQLite collector backend and a Next.js, Tailwind CSS dashboard featuring a Payload Diff Inspector.

Live Production Demo: - 🖥️ Interactive Web Dashboard: https://agent-trace-zeta.vercel.app/ https://agent-trace-zeta.vercel.app/ - ⚙️ FastAPI Swagger Docs: https://agenttrace-api-cdav.onrender.com/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 of 1200.0 - 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 In project root python -m uvicorn main:app --reload --port 8000