HSCodeComp tests whether agents can apply hierarchical professional rules rather than only search the web or complete code. The important point is not the headline alone, but how the announcement changes the practical test for developers, enterprises or policy makers.
What changed
The core facts remain clear: ACL 2026 main conference paper, 632 realistic product entries, 32 categories, 26 customs experts, best current agent 46.8% at 10-digit codes, Marco Agent 65.0%, human experts 95.0%. These details define the scope of the story and keep it grounded beyond launch language.
Why it matters
For readers outside China, the signal is broader than one company update. It shows how AI products are moving from demos toward prices, permissions, hardware limits, energy constraints and measurable deployment results.
What to watch
The next checkpoint is execution: whether the product, platform or policy can hold up in real customer workflows rather than only in benchmark tables or launch-stage examples.
Sources verified: QbitAI, CocoLoop, ACL Anthology 2026.acl-long.937, arXiv 2510.19631, AIDC-AI Marco-DeepResearch GitHub.