Ant Group's confidential-computing unit (Ant Chih Suan / 蚂蚁密算) open-sourced HOP 3.0 at the 2026 Incl...
The AMW Read
Ant's HOP framework advances a known player's approach to the agent-reliability problem with a concrete explore-verify-commit governance mechanism and disclosed reliability/cost metrics, but does not introduce a new top-tier entrant or resolve an open debate.
Ant Group's confidential-computing unit (Ant Chih Suan / 蚂蚁密算) open-sourced HOP 3.0 at the 2026 Inclusion Bund Summit, with chairman Wei Tao unveiling what the unit calls an "agent native language" for building bounded, auditable autonomous agents. HOP debuted at WAIC 2025 as an engineering layer for improving LLM reliability in finance and healthcare; the 3.0 rewrite fuses explicit structured logic — task goals, boundary conditions, verification requirements — with an LLM's own reasoning inside one execution system, and is now published on GitHub.
The framework targets a real production gap: agents that skip steps, drift past intended scope, or leak sensitive data across an unbounded context window during long tasks. HOP 3.0 borrows SQL's commit semantics, splitting agent behavior into an explore phase (reversible, no real-world consequence), independent verification, and a commit gate where irreversible actions — deleting files, dropping a database — are isolated and require explicit clearance. Ant reports that in a spec-driven development workflow using the framework, deliverable completeness and requirement-to-code consistency both hit 100%, average token usage per execution cycle fell about 13%, and failure rates for smaller models dropped roughly 91.7%. Its cited example: a 27-billion-parameter Qwen3 model handling tasks that previously required hundred-billion or trillion-parameter models, because the control burden shifts from the model to the execution engine.
For builders, the pitch is that reliability at production scale no longer has to come from scaling the underlying model — a structured governance layer can let a smaller, cheaper model carry complex, multi-step enterprise work, which matters directly for agent deployment cost in regulated sectors like finance, healthcare, and government. For Ant, open-sourcing HOP ties its confidential-computing business to agent governance infrastructure, positioning data protection and agent trust as one linked stack rather than separate product lines — a distribution angle worth tracking as more enterprises weigh building agent governance in-house versus adopting a framework.