2026-09-11 16:18
At the 2026 Inclusion · Bund Conference, Ant Group Chairman Wei Tao announced that the trusted native intelligent agent HOP 3.0 is officially open source, opening up the relevant technical capabilities of "intelligent agent native language" to developers, enterprises, and industry experts, promoting industrial intelligent agents to move from relying on models to trustworthy execution with clear boundaries, controllable processes, and verifiable results.
During the 2025 World Artificial Intelligence Conference, Ant Computing released and open-source the HOP 1.0 technology framework for the first time, exploring the use of engineering methods to enhance the reliability of large models in professional scenarios such as finance and healthcare. During the 2026 World Artificial Intelligence Conference, HOP will upgrade to version 3.0 and propose the concept of "Agent Native Language", which will further evolve from a trusted application technology framework to a new generation technology system centered on Agent Native Language and aimed at building trusted Native Agents.
Currently, autonomous intelligent agents have demonstrated significant productivity in fields such as software development. However, after entering serious scenarios such as finance, healthcare, and government affairs, they still face three levels of loss of control: human control over AI - the inability to keep up with the generation speed of intelligent agents through auditing and intervention; The loss of control of AI itself - intelligent agents may jump, cross boundaries, or perform erroneous actions, and long-term tasks may fall into a logical quagmire; AI induced data loss of control - sensitive data leakage caused by borderless context.
From the disconnect between Skill and Harness, towards the native language of intelligent agents
From an architectural perspective, the current intelligent agent ecosystem is generally in a fragmented state: task knowledge and operational requirements are written in natural language skills, while true control and scheduling are undertaken by the code Harness. Natural language skills are easy to read and write, but they are difficult to accurately carry complex business flows composed of multi-step, multi branch, multi condition, and multi role collaborations; Although Code Harness can supplement tools, context, and process control, it is mainly aimed at program execution, making it difficult for industry experts to directly participate, and even programmers find it difficult to conduct repeatable audits of execution logic scattered in code, prompts, and runtime states.
In response to this fragmentation, HOP 3.0 proposes an agent native language that integrates explicit structured logic and fuzzy reasoning logic of large models into the same set of expression and execution systems in a concise manner. Explicit structured logic is responsible for clarifying task objectives, boundary conditions, data dependencies, key processes, and verification requirements, while large models make judgments, plans, and explorations within the boundaries.This fusion provides production level reliability for intelligent tasks and fully unleashes the intelligent power of the amplification model within the structural boundaries; On the other hand, transforming the work of intelligent agents into structured objects that are readable, aligned, assessable, and modifiable by industry experts, thereby achieving "locking in goals, guarding boundaries, and opening up paths".
Exploration Verification Submission: Empowering Intelligent Agents to Explore Without Disrupting
It is understood that HOP 3.0 pays tribute to the commit mechanism of SQL transactions, explicitly separating "exploration verification commit" at the language and execution layers. The intelligent agent first explores and experiments in the reversible space, and can only enter the commit submission gateway after the execution results are independently verified; Irreversible operations such as deleting emails, deleting code, and deleting libraries are isolated in the explicit submission phase.
This means that, under the premise of safety and reliability, the system can fully unleash the intelligent exploration ability of the agent: the agent can boldly try and make mistakes, but the exploration experiment itself cannot directly cause destructive consequences to the real system. The exploration path that successfully passes verification can be further solidified into a reusable execution process, allowing an effective exploration to naturally precipitate as an intellectual asset for sustainable reuse and improvement of the organization.
Reduce the pressure of controlling long tasks, and simultaneously enter the production level with reliability and cost
Wei Tao further proposed that HOP 3.0 language can delegate the global goals, step states, branch conditions, context management, and verification requirements of long-term and complex tasks to the language and execution engine, allowing large models to complete reasoning and judgment only within the currently constrained nodes, thereby significantly reducing the requirements of long-term tasks and composite logic for LLM's continuous control capability.
Taking Qwen3.0 27B (approximately 27 billion parameters) in ant computing practice as an example, this model can effectively carry complex tasks that previously required hundreds of B (hundreds of billions of parameters) or even T-level (trillions of parameters) LLMs to undertake. This is not about making ordinary models smarter out of thin air, but about using structured language and execution engines to share the complex task control burden that originally weighed on the model.
As a result, HOP 3.0 not only provides production level reliability, but also allows models and computing power costs to enter the production level usable range, rewriting the cost account for intelligent agents to move from a small number of pilot projects to large-scale production.
According to Ant Group's disclosure, in the practice of complex specification driven R&D processes, after introducing HOP 3.0, the product integrity rate, generation success rate, and consistency between requirements and code all reached 100%, and the average token consumption per execution cycle decreased by about 13%; The failure rate of the ordinary model decreased by about 91.7%, and the effect combined with HOP 3.0 reached or even exceeded that of the strong model without HOP.
Wei Tao stated that open source HOP 3.0 is not only about opening up a set of technologies, but also about allowing developers to participate in the construction of trusted native intelligent agents with experts in risk control, compliance, healthcare, auditing, and other fields. Non programmer experts can integrate business rules, professional judgment, and industry experience into intelligent agent processes; Developers can continuously expand around languages, engines, tools, and industry scenarios, gradually consolidating validated task paths into reusable organizational intellectual assets, reducing the "intellectual islands" where professional abilities remain in a few personal experiences for a long time.
It is reported that the trusted native intelligent agent HOP collaborates with encrypted computing: encrypted computing preserves data, models, and knowledge, while HOP ensures the trusted actions of intelligent agents. The two work together to establish a trusted link between "data model intelligent agent business results". This open source will further gather the power of the industrial ecosystem, promote the formation of a complete chain of "data daring to circulate, models daring to use, and intelligent agents trustworthy", and accelerate the transformation of industrial AI from single point exploration to large-scale application and trustworthy inclusiveness.
It is reported that HOP 3.0 has been officially open sourced on GitHub. Release URL: https://github.
