Instrument agents with frameworks

AI Observability framework integrations capture generations automatically by attaching callbacks or hooks to your agent framework. This eliminates the need to manually instrument each LLM call.

Supported frameworks

Framework	Python	TypeScript	Go	Java
LangChain	Yes	Yes	—	—
LangGraph	Yes	Yes	—	—
OpenAI Agents	Yes	Yes	—	—
LlamaIndex	Yes	Yes	—	—
Google ADK	Yes	Yes	Yes	Yes
Vercel AI SDK	—	Yes	—	—

Set up a Python framework integration

Install the framework-specific package alongside the core SDK:

pip install sigil-sdk sigil-sdk-langchain

Attach the AI Observability callback handler to your framework. For LangChain:

from sigil_sdk import Client, ClientConfig
from sigil_sdk_langchain import SigilLangChainHandler

client = Client(ClientConfig(
    generation_export_endpoint="<SIGIL_ENDPOINT>",
))

handler = SigilLangChainHandler(client)

# Pass the handler to your chain or agent
chain.invoke({"input": "Hello"}, config={"callbacks": [handler]})

client.shutdown()

Each framework integration follows the same pattern: create a handler, pass it to your framework’s callback mechanism, and the integration captures all LLM calls as generations.

Set up a TypeScript framework integration

Import the framework sub-module:

import { SigilClient } from "@grafana/sigil-sdk-js";
import { SigilLangChainHandler } from "@grafana/sigil-sdk-js/langchain";

const client = new SigilClient({
  /* config */
});
const handler = new SigilLangChainHandler(client);

// Pass the handler to your chain or agent
await chain.invoke({ input: "Hello" }, { callbacks: [handler] });

await client.shutdown();

Conversation ID mapping

Framework integrations automatically map conversation IDs from framework context:

If the framework provides a session_id, conversation_id, or group_id, the integration uses it.
If a thread_id is available (LangGraph, OpenAI Agents), the integration uses it.
Otherwise, the integration generates a deterministic ID from the framework run context.

Capture workflow steps

Some framework integrations can capture workflow steps in addition to LLM generations. Workflow steps represent non-LLM execution nodes, for example, routing, planning, retrieval, or tool orchestration.

Enable workflow step capture when you want AI Observability to show the agentic workflow graph around the LLM calls. The integration links workflow steps to the generations they contain and records parent step IDs so the conversation graph can show the execution order.