auto-chat

Repository for creating agent-agent conversations.

There are two ways to run this repository. The first way, is to have the handler agent (GPT-5) come up with the scenario. This will include generating a user profile, a set of user interests and a goal that the handler agent is trying to accomplish.

The default behavior is for the agent to continue to the conversation until its goal has been accomplished. We also provide a flag --max_turns to control how long the conversation can continue (in the case that the model is unable to accomplish the agent's goal).

Another useful flag is max_tool_calls which controls the number of consecutive tool calls that can be made without the agent responding. The default is 5, but that might be too low for some scenarios that require looking at the user's open tabs and reading the pages.

Setup

This package is mostly python based, but to replicate the Firefox browser, we use readability.js for webpage parsing. In addition, we use playwright as a headless browser to navigation. To setup your environment, run the following commands:

1. install node dependencies (readability.js)

npm install

2. install python dependencies

uv sync

3. install playwrite browsers

uv run playwright install

Example usage -- generated scenario

uv run python start_chat.py --model together.ai:Qwen/Qwen3-235B-A22B-Instruct-2507-tput\
                            --eval_model_id openai:gpt-5\
                            --max_turns 5

If you wish to create a custom scenario, you can add a .yaml file to the conversations directory. Sample scenarios are provided. The easiest thing to do would be to use one of these as a template and update accordingly.

Example usage -- custom scenario

uv run python start_chat.py --model together.ai:Qwen/Qwen3-235B-A22B-Instruct-2507-tput\
                            --eval_model_id openai:gpt-5\
                            --conversation_file conversations/haunted_dinner_menu.yaml\
                            --max_turns 5