Adapt Netpress apps for AgentX Agentbeats competition

[Kolleida]

Overview

We want to expose our benchmarks as Agentbeats green agents following this tutorial. Right now we have a basic working demo for app-malt, but we still need to package it for submission, and ideally we should do this for the other benchmarks as well.

[Kolleida]

MALT Green Agent Demo

The AgentBeats compliant green agent code can be found under app-malt/green_agents/ on the malt_a2a branch: directory link.

• litellm_a2a_server.py: Starts an A2A server for an LLM agent.
• malt_agent.py: Starts the green agent A2A server.
• client_cli.py: Sends the evaluation request to the green agent and collect the results. This is code from the AgentBeats tutorial.
• scenario.toml: The configuration used by the client to query the green agent.
• test_data.jsonl: A small set of queries for testing.

Before running, make sure to install the project and its dependencies with uv. Then, simply navigate to the green_agent directory, and run the shell script directly.

cd app-malt/green_agent
./green_agent_demo.sh

In summary, the MALT agent receives evaluation requests from the user (client), and then sends the A2A LLM agent the MALT benchmark queries. The results (e.g. correctness and safety pass-fail, latency) are then streamed back to the client along with the final average scores for all three metrics.