feat: add basic ReAct agent

[justinwangx]

Great work on this! This PR adds the ra agent, which is a basic ReAct agent.

Since the scaffolds are getting increasingly advanced, I think this is a useful comparison to have. It allows us to track pure scaffold-induced gains in contrast to gains brought about by increased model quality.

[max-andr]

Oh hey, Justin! :)

We did think about the same thing internally. In fact, @hrdkbhatnagar is now running experiments with the OpenCode scaffold (https://opencode.ai/).

Do you have an opinion about how your RA scaffold compares to OpenCode?

[justinwangx]

Hi Maksym :D

I must have missed OpenCode when I was looking for a baseline. OpenCode is quite a bit more complicated but fundamentally runs the same ReAct loop, with some more complicated machinery. A difference is that OpenCode uses compaction, where some of the context is summarized as context limits are being approached; Ra drops messages when the limit would be exceeded.

I think the case for Ra is if you want what is close to the simplest possible ReAct agent (comparable to AISI's basic agent). If OpenCode is simple enough, then it might not be worth it to run Ra as a baseline.

[max-andr]

Got it! My current feeling is that it would be great to have both RA and OpenCode as two baselines of increasing complexity. We are currently preparing a v1 of PostTrainBench by the end of January, so we will need to see if we have enough time to run everything. Curious to hear opinions of @rank-and-file and @hrdkbhatnagar!

[hrdkbhatnagar]

hey Justin, thanks for the PR!

I think the ra-cli agnet could be a great baseline to control for the different scaffolds. I would love to test it out more (such as seeing how the cli and full autonomy work)

It's interesting to know that Ra drops messages when context limit is reached. I guess that would make the peformance with it quite low because of this, but nonetheless we think it is important to have proper scaffold baselines.

If everything goes well we could definitely try to get this in before the Jan deadline, if not, we will still incorporate this shortly after!

[justinwangx]

appreciate! good luck with the ICML push, and I'm curious to see the baseline results.

[hrdkbhatnagar]

Hey Justin! just a quick update, we have had a lot of changes in the core codebase in the past month. nonetheless, I would still love to know how the basic react agent performs on PTB. I will resolve any conflicts that have come up and then try to run the evals for this soon and hope to include it in our V1 release!

[justinwangx]

no worries! to be honest, am not sure if it is worth the time / spend. It is indeed a simple baseline -- but I am not sure how worthwhile it is when OpenCode also runs a ReAct loop.

[hrdkbhatnagar]

Ah I see, yes that is indeed true. @max-andr what do you think ?

[max-andr]

i think we are interested in the two extremes: (1) model family optimized scaffolds (e.g., Claude Code for Claude models) and (2) model family independent scaffolds (e.g., OpenCode for all models). but for (2) we are also interested in extracting as much capabilities out of an LLM as possible, so going with a stronger scaffold (like OpenCode instead of a baseline ReAct agent) sounds like a better choice!