how do i let the agent respond in between tool calls?

[gutbash]

for example chatgpt.com can dynamically...

1. respond to the user
2. use a tool
3. respond to the user again
4. use another tool
5. return a final response.

how do i do this with langchain? currently i am streaming async events from AgentExecutor with tools and it has never behaved like this.

[dosubot[bot]]

Hi @gutbash! I'm Dosu and I’m helping the langchain team.

LangChain's AgentExecutor does support streaming and iterating over steps, which lets you surface intermediate reasoning and tool actions as they happen—similar to how chatgpt.com alternates between responding and tool use. This is done via the stream and astream methods, which yield each step (including tool calls and responses) as they occur, using the AgentExecutorIterator under the hood. However, the agent loop is still bounded per run: you can't inject new user input mid-run, but you can stream each intermediate action and response as they happen within a single run, allowing for dynamic, multi-step interactions example.

If you want even more flexibility (like pausing, branching, or richer control flow), LangChain now recommends using langgraph, which is designed for advanced agent workflows and dynamic interactions.

If your current AgentExecutor setup isn't yielding intermediate responses as expected, make sure you're using the stream/astream interface and not just waiting for the final output. Also, if you're using OpenAI tools, switch from create_openai_functions_agent to create_openai_tools_agent to avoid repeated tool calls and get the latest tool-calling behavior reference.

Let me know if you want a code example or more details on migrating to langgraph!

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}