Overview
AgentOmnia studies full-scenario agentic scaling: how to improve an agent across consumer, business, and employee applications rather than optimizing for a small collection of isolated tool-use benchmarks.
Method
The framework connects a Domain × Capability × Atomic Difficulty taxonomy with executable environment and task synthesis, supervised fine-tuning, online agentic reinforcement learning, and evaluation-driven iteration. Programs, solvers, and verifiers provide reliable signals for difficult tasks, while evaluation failures can be converted into Product Requirement Documents for subsequent data construction.
Results
Based on Qwen3-30B-A3B-Thinking-2507, AgentOmnia raises the OmniaBench challenging-subset pass rate from 9.16% to 37.11% and the macro-average across OmniaBench, τ²-Bench, DeepPlanning, and VitaBench from 22.86% to 41.69%. Improvements cover 76 of 90 level-one domains and all evaluated capability and atomic-difficulty dimensions.
Perspective
From my perspective, the most valuable part of AgentOmnia is not a single benchmark score, but the shared coordinate system connecting task definition, data synthesis, post-training, and evaluation. It turns model failures into actionable training requirements. The current results demonstrate broad improvement at a fixed model scale, while stronger base models remain ahead on some tasks and PRD-guided self-evolution still requires validation over more data and training rounds.