Adobe and UC San Diego build agent that routes image edits like a professional

EditAgent's architecture isn't novel in concept (tool composition has been explored), but the 10x computational cost is the actual constraint being surfaced here. Adobe is essentially asking whether professional-grade creative work requires accepting massive overhead, or whether that overhead signals the approach won't scale to production.

This connects directly to the broader agent infrastructure race covered in recent weeks. Google's Android framework (late October) positioned agents as pluggable, interoperable components. OpenAI's dots initiative (late October) pushed toward always-on persistence. Apple embedded agents into productivity tools (late October). EditAgent represents the inverse move: instead of building general-purpose agents that route to tools, Adobe is building task-specific tool orchestration that uses LLMs as evaluators. The efficiency penalty suggests that domain-specific agent design (creative editing) may not follow the same cost curve as general coding or writing agents. If specialized workflows demand 10x overhead while general-purpose agents don't, that's a fundamental split in how the industry will architect AI assistance.

If Adobe ships EditAgent as a production feature within Photoshop in the next 12 months without substantially reducing the computational overhead, that signals the company believes the cost is acceptable for professional users. If they shelve it or pivot to a lighter-weight approach, watch whether they publish findings on why the overhead couldn't be justified, which would inform whether other creative tool makers (Figma, DaVinci Resolve) pursue similar architectures or avoid them entirely.