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

> Source: <https://github.com/mohitkumar188/AgentTrace>
> Published: 2026-09-20 21:24:17+00:00

**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
```