AIBRIEF
Borrowed from database transaction logic, the gate keeps agents in a reversible rehearsal space until a separate check passes, then allows the irreversible action to land on the real system.
Ant Group's confidential-computing arm, 蚂蚁密算 (Ant MiSuan), open-sourced HOP 3.0 on Friday, an agent framework that borrows one idea from databases: a SQL-style "commit" gate.
Today's serious agents drift off scripts, delete the wrong file, or leak prompt data into a log. HOP 3.0 answers with a three-step loop the company calls explore, verify, commit. The agent explores in a reversible rehearsal space, an independent check has to pass, and only then does an irreversible action like deleting an email or dropping a database table reach the real system. The rules live in HopSpec, a structured-markdown task spec, and the runtime lives in HopJIT, the engine that enforces them.
Chairman 韦韬 framed the release as "agent-native language" at Shanghai's Inclusion Bund conference. The company reports that, in its own R&D workflow, HOP 3.0 hit 100% spec completeness, cut per-cycle token use about 13%, and dropped an ordinary model's failure rate roughly 91.7%. Those numbers are vendor-disclosed on a single internal workflow, not third-party benchmarks; the repo and Chinese tech coverage trace back to the same launch announcement.
The pitch: a 27-billion-parameter Qwen model paired with HOP 3.0, the company says, can handle the complex tasks that previously required models hundreds of times larger. Independent tests will tell.