Draft ae-agent

[xuafeng]

[Bogdan, disguised as Xuan making changes :-) ]

I'm using this umbrella issue to track the design and implementation of the agent.

Issues, design decisions, discussion:

1. Redesign master plan: https://github.com/sys-intelligence/artifact-agent/issues/5
2. Important agent tools to add: https://github.com/sys-intelligence/artifact-agent/issues/1

---

moved previous suggestions in a comment below

[xuafeng]

Suggestion 1
The agent in CLI format (run on any env, specific hw, docker)

• input artifact -> output playable env (local/docker) and reproduced results.

Then, researchers/students can play with it, gain deeper understanding and may propose better solutions.
It is a good agent beyond artifact evaluation.

Suggestion 2
Repos w/ artifact README.md are a good source to build runnable env.

Then, the model can play w/ the env, to have better system (basically better score on the reported criteria, easily verify via the table and figure, get the reward)

Good data for RL, suitable for optimization, debug, xx tasks.

[bastoica]

@xuafeng @Couen We should probably have an offline discussion about the requirements. I sketched out a design, and your input would be super valuable.

[bastoica]

I'll use this issue as an umbrella to track agent-related issues, documentation, design discussions, etc.