Unified multimodal large language models (MLLMs) and multi‑agent systems have advanced visual generation, yet three limitations persist: (1) existing approaches distill task‑specific experience with limited generalizability; (2) reflection is deferred until after task completion; (3) knowledge is acquired only when downstream demands arise.
OmniHarness addresses these issues by abstracting verified executions into symbolic policies. A symbolic policy captures shared generation procedures and applicability conditions while stripping instance‑specific inputs, thus forming reusable templates for families of visual‑generation tasks.
The harness component instantiates, adapts, and composes these policies for new tasks. Intermediate verification monitors execution, guides policy refinement, and enables failure recovery. Moreover, OmniHarness conducts self‑directed inquiry, generating practice tasks near its capability limits before downstream objectives are defined, allowing continuous feedback‑driven improvement while keeping model parameters frozen.
Experiments span six benchmarks, three MLLM backbones, and three visual‑agent frameworks. OmniHarness attains a 95.0% resolve rate on Creative tasks in ComfyBench, surpassing the strongest baseline by 27.5 percentage points, and demonstrates continual capability expansion. Frozen policy snapshots can be plugged into existing visual agents to boost performance without retraining.
Review