How do callbacks work in LangChain?

Callbacks in LangChain are hooks that fire at specific lifecycle events during chain, model, and agent execution. You implement a BaseCallbackHandler subclass and override only the methods you care about. Each method receives context about what just happened — which model was called, what the prompt was, what the response was, and how long it took.

Key callback methods (all have async equivalents prefixed with a):

• on_llm_start(serialized, prompts) — fired before an LLM call
• on_llm_end(response) — fired after an LLM call completes
• on_chain_start(serialized, inputs) — fired when a chain begins
• on_chain_end(outputs) — fired when a chain finishes
• on_tool_start(serialized, input_str) — fired before a tool executes
• on_tool_end(output) — fired after a tool returns
• on_agent_action(action) — fired each time an agent decides to use a tool

from langchain_core.callbacks import BaseCallbackHandler

class TokenLogger(BaseCallbackHandler):
    def on_llm_end(self, response, **kwargs):
        usage = response.llm_output.get('token_usage', {})
        print(f"Tokens used: {usage}")

chain.invoke({"input": "hello"}, config={"callbacks": [TokenLogger()]})

Callbacks can be attached per-invocation via config={"callbacks": [...]}, per-component via constructor arguments, or globally with set_global_handler(). LangSmith tracing itself is implemented as a callback handler.