## The problem
Agent 1 hallucinates. Agent 2 builds on it. Agent 3 executes it. All three returned HTTP 200. Nobody knew.
In multi-agent AI pipelines, errors propagate silently. Existing tools like LangSmith and Langfuse log what happened after. AgentTrace stops it before.
## What it does
AgentTrace wraps any agent in 2 lines and acts as a circuit breaker:
- ⚡ One agent blocked = entire pipeline stops — downstream agents never run
- 🛡️ 13 built-in rules — hallucination detection, PII leakage (SSN, credit cards), prompt injection, OWASP LLM Top 10, EU AI Act compliance
- 📊 Real-time dashboard — every trace, violation, audit ID, risk score
- 💾 Zero cloud — append-only NDJSON, everything stays local
## Works with your stack
```ts
const guard = new AgentTrace({ rules: ['block_hallucination', 'block_pii_leakage'] });
const safeAgent = guard.wrap(myLangChainAgent);
