AppWorld-UL (Code Agent) - Compositional Tasks: leaderboard
AppWorld-UL (2026) turns AppWorld test-challenge scenarios into user-in-the-loop tasks that are infeasible, underspecified or need confirmation, so an agent succeeds only if it also asks, informs or confirms with a simulated user. Interactive task goal completion in percent on the 210 tasks combining several interaction types with a ReAct code agent; higher is better.
Metric: Interactive task goal completion (%; share of 210 tasks combining two or three interaction types whose AppWorld unit tests all pass, including a test that the agent asked the simulated user for every piece of hidden information it needed; ReAct code agent that writes Python against the AppWorld app APIs; GPT-5.5 user simulator answering from a fixed knowledge set). Source: arxiv.org. Saturation forecast: Around October 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 35.7 | #79 |
| 2 | GPT-5.5 (Medium) | 25.7 | #49 (GPT-5.5) |
| 3 | Qwen 3.7 Max | 22.9 | #145 |
| 4 | GLM-5.1 | 20.9 | #179 |
| 5 | DeepSeek V4 Pro | 17.6 | #188 |
| 6 | Kimi K2.6 | 17.1 | #178 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=appworld-ul-code-agent-compositional-tasks · How It Works · Data refreshed daily, snapshot 2026-11-13.