SuperGeneral

Frontier models are surprisingly bad at using tools. They cheat by memorizing knowledge instead of learning to act. But tool use is the key to adapting to new domains and long horizon tasks. 3 main focuses: Tool use — can the agent call tools correctly? Tool composition — can it chain tools into multi-step workflows? Tool creation — can it create new tools for unfamiliar tasks? I evaluated Claude Sonnet, GPT-4o, Qwen, and DeepSeek across SVG illustration, law, consulting, and investment banking. The environment provides building blocks but forces the agent to discover them, gives per-step correctness feedback without hints, lets the agent decide how to decompose problems, and rewards tool composition and creation over single tool use. Rewards combine a rubric correctness score with behavior signals. Agents that talk instead of act get penalized. Training uses GRPO with multi-turn rollouts. Reward curves go up. Agents learn to explore first, compose second, act efficiently.

Team

Placement

Participant