Skip to content
Bud's Wiki

AgentOmnia: Scaling Agentic Models for Full-Scenario Applications

TL;DR

Section titled “TL;DR”

AgentOmnia is a full-stack agentic post-training framework from Huawei/Zhejiang that organizes agent training around a Domain × Capability × Atomic-Difficulty taxonomy spanning To-Consumer, To-Business, and To-Employee applications. It synthesizes 5,018 stateful environments with 255,375 tools and 52,361 tasks using bidirectional environment-task synthesis plus tool-dependency, program-structured, and solver-based pipelines, then combines SFT + online agentic RL + a rollback curriculum. Starting from Qwen3-30B-A3B-Thinking-2507, it lifts the challenging-subset pass rate from 9.16% to 37.11% and beats Qwen3-235B-A22B-Thinking-2507 on all four evaluated benchmarks. Evaluation failures are converted into Product Requirement Documents (PRDs) that drive targeted self-evolution — an early loop toward automated agent improvement.

Key claims

Section titled “Key claims”
  • The Domain × Capability × Atomic Difficulty taxonomy is the alignment substrate that unifies task definition, data synthesis, post-training, and evaluation into one pipeline instead of ad-hoc per-benchmark tuning [§1, §3].
  • Bidirectional environment-task synthesis — combined with tool-dependency, program-structured, and solver-based pipelines — produces 5,018 stateful environments, 255,375 tools, and 52,361 tasks with executable correctness signals from programs, solvers, and verifiers [§4].
  • Post-training combines SFT, online agentic RL, and a rollback curriculum; starting from Qwen3-30B-A3B-Thinking-2507, OmniaBench challenging-subset pass rate rises from 9.16% → 37.11% and the macro-average across OmniaBench, τ²-Bench, DeepPlanning, and VitaBench rises from 22.86% → 41.69% [Abstract].
  • Under a unified evaluation protocol, AgentOmnia leads the evaluated agentic post-trained baselines on OmniaBench and surpasses Qwen3-235B-A22B-Thinking-2507 on all four benchmarks — i.e. a 30B activation-parity model beats a 235B one via better environments + curriculum [Abstract].
  • Gains are broad, not category-specific: improvements across 3 application splits, 10 capability dimensions, 8 atomic-difficulty factors, and 76 of 90 level-1 domains [Abstract].
  • Evaluation failures are converted into Product Requirement Documents (PRDs) that drive a self-evolution loop; a one-round study provides initial evidence, motivating larger-scale validation [Abstract].

Method

Section titled “Method”

AgentOmnia treats agentic scaling as a coordinated pipeline rather than a training recipe. A Domain × Capability × Atomic Difficulty taxonomy is fixed up front and used to specify the target task space across ToC/ToB/ToE. Environment and task synthesis is bidirectional: environments seed tasks that require particular tool compositions, and desired capability targets seed the environments that instantiate them. Three complementary pipelines produce the training data — tool-dependency (graphs of which tools depend on which state), program-structured (task recipes expressed as programs), and solver-based (reference solvers whose traces become supervision). The result is 5,018 stateful environments (255,375 tools, 52,361 tasks) where correctness is checked automatically by programs, solvers, and verifiers.

Post-training layers SFT (behavior cloning from solver traces), online agentic RL (with those same programs/verifiers as reward signal), and a rollback curriculum that revisits earlier tasks when later stages regress. The evaluation loop is closed: benchmark failures are packaged into Product Requirement Documents (PRDs) that describe what to fix, and the pipeline consumes those PRDs as inputs to the next data-synthesis round — an initial self-evolution study.

Results

Section titled “Results”

Starting checkpoint is Qwen3-30B-A3B-Thinking-2507. Headline numbers [Abstract]:

  • OmniaBench challenging subset: 9.16% → 37.11% pass rate.
  • Macro-average across OmniaBench + τ²-Bench + DeepPlanning + VitaBench: 22.86% → 41.69%.
  • Beats every evaluated agentic post-trained baseline on OmniaBench under the unified protocol.
  • Surpasses Qwen3-235B-A22B-Thinking-2507 on all four benchmarks and exceeds Qwen3.5-35B-A3B on the macro-average — i.e. a 30B/A3B checkpoint after AgentOmnia post-training beats the 235B/A22B checkpoint of the same family.
  • Improvements span 3 application splits, 10 capability dimensions, 8 atomic-difficulty factors, and 76 of 90 level-1 domains.

Why it’s interesting

Section titled “Why it’s interesting”

AgentOmnia sits in the same design space as AREX: Towards a Recursively Self-Improving Agent for Deep Research and Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem — full-stack systems that co-design env synthesis + curriculum + RL rather than iterating on the RL algorithm alone — but pushes further on breadth (255K tools, ToC/ToB/ToE cross-cut) and adds a PRD-driven self-evolution loop, which is a concrete counterpart to the recursive-self-improvement thread (2607.14pd-tao0-vla-hierarchical-vla-foundation-2026-07 is unrelated; compare with AIDE²: First Evidence of Recursive Self-Improvement). It also complements Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning and Toolathlon-GYM: Large-Scale Long-Horizon Environments for Tool-Use Agents, both of which stop at environment synthesis; AgentOmnia’s contribution is that the environments are wired straight into a Difficulty × Capability curriculum with rollback and PRD-derived data patches. The 30B-beats-235B result also complements the MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training line: post-training data quality + curriculum can substitute for order-of-magnitude parameter scale on agentic tasks.

See also

Section titled “See also”