Dan Austin

@AlexGDimakis Thanks! Yes exactly, the orchestrator is the advisor trained with GRPO. The subagents are also open models (Qwen3-Coder), however they are left alone during training and used purely as tool calls. Cost implications prevented using closed models / GLM-4.6

@DAu5tin cool work ! so if I understand, the advisor in your architecture is the orchestrator? Is the orchestrator trained with GRPO while the other agents are closed models?

I scaled coding-Agent RL to 32x H100s. Achieving 160% improvement on Stanford's TerminalBench, and it was fun! ⚡️🤓

Also thank you to @AlexGDimakis who sparked the idea to train the Orchestrator model (in multi-agent setup), instead of a single agent doing everything. When we discussed the now released paper "How to Train Your Advisor". Thanks Alex! arxiv.org/pdf/2510.02453

Thank you to @zjasper and the team at @hyperbolic_labs for providing such a great GPU cloud service, and for always being available for any issues or queries! 10/10!

Me on Stanford's TerminalBench leaderboard ahead of Claude Code! 🤓 Forgot to post this image before

Lots more detail in here: github.com/Danau5tin/mult…

Here is how it works at a high level! Open source repo below (2/3)

Weekend experiment accidentally beat Claude Code on Stanford's TerminalBench (#12) - turns out orchestrating multiple specialised AI agents with shared memory works better than I expected 🤓 Open sourced! (1/3)

All open source with a 208 ⭐️'s github.com/Danau5tin/term…

My name on Stanford's terminal bench leaderboard🤓 Would love to see how far it would go with some RL compute⚡️

@_manan2005 Hard to say without knowing the details, I guess a good general rule is to make sure the task you are trying to solve for is one which takes a long time, and then within RL, incentivise the agent (via reward) for not trying to finish too early.

@DanAiTuning Can you give some tips on how to make an agent running for hours I am trying to train one but the synthetic. conversation ends up in like 20 to 25 max rounds

Just tested my long-horizon terminal agent RL training on 32x H100s! 🚀 Too GPU poor to actually train though 😅 Happily though, my untrained agent hit #19 on terminal bench!

@zjasper @hyperbolic_labs Yes thank you for such a great platform!

@DanAiTuning Excited to see that you trained this agent using @hyperbolic_labs!

Everything is open source (agent, data, training code) github.com/Danau5tin/term…

Also, a huge thank you for the amazing work on TerminalBench, which gave me the inspiration to push an AI agent to the limits and be able evaluate it. 🧗🏔️ Thanks to @Mike_A_Merrill, @alexgshaw and all the contributors to this amazing benchmark.

Built on the amazing rLLM package - which is a great dev experience for multi-turn RL. Their framework just works and handled everything I threw at it 🙏 Thanks to @michaelzluo, @sijun_tan, @_royh021, and all of the amazing team at @Agentica_