Hitting a million brick walls with multi-turn RL training isn't fun, so I thought I would try something new to climb Stanford's leaderboard for now! So this weekend I was just tinkering with multi-agent systems and... somehow ended up beating Claude Code on Stanford's TerminalBench leaderboard (#12)! Genuinely didn't expect this - started as a fun experiment and ended up with something that works surprisingly well.

*What I did:*

- Built a multi-agent AI system with three specialised agents:

- Orchestrator: The brain - never touches code, just delegates and coordinates

- Explorer agents: Read & run only investigators that gather intel

- Coder agents: The ones who actually implement stuff

- Created a "Context Store" which can be thought of as persistent memory that lets agents share their discoveries.

- Tested on TerminalBench with both Claude Sonnet-4 and Qwen3-Coder-480B.

*Key results:*

- Orchestrator + Sonnet-4: 36.0% success rate (#12 on leaderboard, ahead of Claude Code!)

- Orchestrator + Qwen-3-Coder: 19.25% success rate

- Sonnet-4 consumed 93.2M tokens vs Qwen's 14.7M tokens to compete all tasks!

- The orchestrator's explicit task delegation + intelligent context sharing between subagents seems to be the secret sauce

*(Kind of) Technical details:*

- The orchestrator can't read/write code directly - this forces proper delegation patterns and strategic planning

- Each agent gets precise instructions about what "knowledge artifacts" to return, these artifacts are then stored, and can be provided to future subagents upon launch.

- Adaptive trust calibration: simple tasks = high autonomy, complex tasks = iterative decomposition

- Each agent has its own set of tools it can use.

*More details:*

My Github repo has all the code, system messages, and way more technical details if you're interested! (Github handle is danau5tin).

Thanks for reading!

Dan

Curious how cheaper models would fare. Have you thought about testing gpt5-mini and similar? And even the very small qwen-coder-30b.