Adobe Research and UC San Diego developed EditAgent, an AI system that decomposes image editing into planned steps, pairing a general-purpose editor (Flux-2-Klein-9B or Qwen-Image-Edit-2511) with specialized tools for sharpening, rotation, and zooming. On a 600-edit benchmark blind-rated by five professionals, the base editor failed 42.7% of the time; specialized tools resolved roughly two-thirds of those failures. Parallel variant generation from the original image was preferred by professionals in 78.5% of comparisons versus iterative refinement, which accumulates errors. EditAgent uses a standard LLM (e.g., Qwen3-8B-27B) as a judge, fed verification records containing close-up crops, change maps, and measured brightness/color statistics for each variant. With this data, seven tested LLMs outperformed fine-tuned evaluators like EditScore-72B; Qwen3-8B-27B selected a usable result 81.3% of the time versus 74.8% for EditScore-72B. EditAgent also detects conflicts where later edits undo earlier ones. On the EdiVal benchmark, the Qwen variant surpassed commercial models Seedream 4.0 and GPT-Image-1.5 by its third editing round, though EditAgent requires ~10x the compute of a single editor. Verified outputs generated ~26,000 training samples; editors fine-tuned on this data improved across three benchmarks, whereas FLUX self-generated data degraded FLUX performance. Black Forest Labs has since released Flux 3 Image Klein (5.6B parameters), which outperforms Flux 2 Klein and Flux 2 Pro in internal tests.