Edition November 2, 2026

Adobe tests how well AI agents use editing tools like a Photoshop professional

EditAgent selects a suitable tool for each task, generates alternatives with different methods in parallel, and has a standard language model compare them together instead of repeatedly retrying one editor.

With verification records, all seven tested language models outperformed every model fine-tuned on human ratings; Qwen3.8-27B selected a usable result 81.3 percent of the time, compared with 74.8 percent for EditScore-72B. EditAgent requires about ten times the compute of a single editor, and its performance depends on the tools connected to it; on the EdiVal benchmark, the Qwen variant narrowly surpassed the commercial models Seedream 4.0 and GPT-Image-1.5 in its third editing round.

This summary is based on the available full source text. Summary prepared with AI assistance. The original source is authoritative.