Framework integration¶

LoopLens is not a standalone app you rebuild your agent inside. You add it to the agent you already have. There are a few ways to do that, from zero-instrumentation to full manual control.

graph LR
    A[Your agent] -->|OTel exporter| B[/v1/traces/]
    A -->|callback handler| C[LangGraph adapter]
    A -->|trace / event| D[Manual SDK]
    B --> E[(LoopLens server)]
    C --> E
    D --> E
    E --> F[Detectors + metrics + live UI]

Pick an approach¶

1. OpenTelemetry — universal, zero code¶

Most agent frameworks — LangChain/LangGraph, LlamaIndex, CrewAI, AutoGen, the OpenAI Agents SDK, and more — can emit OpenTelemetry spans via the OpenInference or OpenLLMetry instrumentations. Point their OTLP exporter at LoopLens and the spans become runs, events, and loop warnings with no LoopLens-specific code in your agent.

→ OpenTelemetry ingestion

2. LangGraph adapter — tight, in-process¶

A single LangChain callback handler that captures every node's LLM and tool calls. One line in your run config.

→ LangGraph adapter

3. OpenAI Agents SDK adapter — native handoffs & guardrails¶

A native TracingProcessor that captures the SDK's typed spans, including handoffs and guardrail trips that the generic OTel path only approximates.

→ OpenAI Agents SDK adapter

4. CrewAI adapter — native crew delegation¶

A BaseEventListener that captures crew events and emits a handoff when control moves to a different agent — so a stuck crew trips handoff_bounce.

→ CrewAI adapter

5. Manual SDK — full control¶

Call trace() / event() / @observe yourself. Best for hand-rolled loops or when you want to emit exactly the events you care about.

→ Manual SDK

Which should I use?¶

You have…	Use
Any framework with an OpenInference/OpenLLMetry instrumentor	OpenTelemetry
A LangGraph or LangChain app	LangGraph adapter (or OpenTelemetry)
An OpenAI Agents SDK app	OpenAI Agents adapter (native handoffs/guardrails)
A CrewAI crew	CrewAI adapter (native delegation handoffs)
A hand-rolled loop, or a framework with no instrumentor	Manual SDK

All of these funnel through the same ingestion pipeline, so the loop detectors, health scoring, and live UI work identically no matter how the events arrived.