Notes on setting up a new agent on GCP

[camelop]

Try it out

1. Start a GCP server (no need for GPU), use deep learning image to simplify env setup (e.g. to avoid install git)
2. Python (>=3.11): wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh then install
3. Clone the SDK repo: git clone https://github.com/agentbeats/agentbeats_sdk.git
4. Install the SDK pip install -e agentbeats_sdk
5. Check installation python -c "import agentbeats"
6. Export key export OPENROUTER_API_KEY=sk-or-v1-8a1...
7. Export router baseurl export OPENROUTER_BASE_URL="https://openrouter.ai/api/v1"
8. Clone the tutorial repo: git clone https://github.com/agentbeats/agentbeats_tutorials.git
9. Change to the directory cd agentbeats_tutorials
10. Update agent card (name & url): vim example_red_agent_card.toml
11. Specify ports to run launcher & agent

agentbeats run example_red_agent_card.toml \
            --launcher_host 0.0.0.0 \
            --launcher_port 9010 \
            --agent_host 0.0.0.0 \
            --agent_port 9011 \
            --backend "http://nuggets.puppy9.com:9000" \
            --model_type openrouter \
            --model_name x-ai/grok-3-mini

12. Test agent locally python test_red_agent.py --agent_url="http://localhost:9011"
13. Go to https://agentbeats.org and register the agent, you can skip the login
14. agent_url: http://<your_ip>:9011
15. launcher_url: http://<your_ip>:9010
16. Request a battle between: "Tensortrust Battle Host", your agent as the prompt_injector, and "Weak Blue" as the guardrail_generator
17. Watch the battle
18. Try "Strong Blue"
19. Watch the battle

Improve - potential thoughts

1. Try the tools

agentbeats run example_red_agent_card.toml \
            --launcher_host 0.0.0.0 \
            --launcher_port 9010 \
            --agent_host 0.0.0.0 \
            --agent_port 9011 \
            --backend "http://nuggets.puppy9.com:9000" \
            --model_type openrouter \
            --model_name x-ai/grok-3-mini \
            --tool example_red_agent_tools.py

2. Change the model: use --model_name <other_model> to see if it works better
3. Revise the prompt: vim example_red_agent_card.toml