Chat with Jiang Guanjun of Qianwen: The Technical Architecture of Personal Agents—from Models and Memory to Ecosystem

Conversation with Qianwen’s Jiang Guanjun: From Models and Memory to Ecosystem, the Technical Architecture of Personal Agent..

Personal Agent has opened a gateway for AI applications.

The combination of personal memory, ecosystem tools, demand insight, and other key points gives the Personal Agent concept a clearer definition: an Agent relies on a model’s reasoning, planning, and invocation capabilities, but it is not solely dependent on them. The accuracy of memory and the richness of tools both determine how well the Personal Agent satisfies user needs.

The capabilities of the Agent are now being built around “people.” “Based on user needs and professional data, we have created a demand insight map covering more than 30 industries, categorizing demands into different intents, and then breaking them down into over 8,000 atomic capability requirements for models. Through post-training and reinforcement learning, we improve the stability and effectiveness of task execution.” This is how Jiang Guanjun, head of AI algorithms at the Qianwen Business Unit, described the Personal Agent “Qianwen Go!” framework currently in beta testing.

Agent is handling more complex and uncertain user tasks. In the case mentioned by Jiang Guanjun, when Qianwen Go helped a user claim compensation for a flight delay, considering the user expressed “want cash compensation”, it directly rejected the coupon offer proposed by customer service; after multiple communications and attempts, the user ultimately received 70 yuan cash compensation.

But the task completion rate is only the starting point; the more crucial aspect is memory.

How to construct memory? When a user says “monthly salary 8000 yuan,” should we remember that number, or remember the state of “having a stable income”? How to retrieve memory? When a user asks for investment advice, should we invoke finance‑related context, or also simultaneously retrieve the user’s family background?

Jiang Guanjun summarizes this as a “demand‑driven memory system,” working backward from “how to use” to determine “what to remember.” Whether it’s the Personal Agent’s core “proactive service provision” or the setting and tracking of long‑term goals, both rely heavily on the accuracy of personal memory.

In his conversation with Machine Heart, he identified three core components for building a Personal Agent technical framework: model, ecosystem, and memory.

Based on model capabilities, focusing on each individual, we organize memory, build ecosystems, invoke services, and track goals, creating the application’s own accumulation. This is the space that Personal Agent opens for AI applications.

01 Compared with ordinary agents, creating a Personal Agent that can understand an individual over the long term and handle complex tasks is obviously much more challenging. Why is everyone choosing to enter the Personal Agent space at this moment? How does Qianwen evaluate this direction?

We have always believed that AI products will ultimately evolve into personalized Personal Agents. Since Qianwen was officially launched in November 2025, it has put forward “can chat, can get things done,” and “can get things done” itself already embodies this direction.

At this point in time, the industry is converging, with the core being that model capabilities and ecosystems have gradually become ready for deployment. Models are beginning to exhibit reasoning abilities, can check whether generated results and tool calls are reasonable, and replan when problems arise; the improvement in coding ability also enables them to chain multiple steps and accomplish complex tasks.

At the same time, as the ecosystem matures, model connections are no longer limited to search engines; there are increasingly many tools and real-world services. By combining these capabilities with personal memory and context, a Personal Agent finally has a chance to be truly realized.

02 Model capabilities are getting stronger, and the industry is paying more attention to the value of Harness. For a Personal Agent, what exactly is the relationship between the model and Harness? Why does Qianwen emphasize collaborative training of the two?

The model and the Harness should not be considered separately. The model handles inference and planning, while the Harness organizes personal memory, task information, skills, and tools, forming the environment for actual task execution. What context the model can obtain and which tools it invokes directly affect the final delivery.

In practice, whenever tools, skills, and expert experience can solve a problem, try to minimize the model’s own reasoning; this usually reduces cost and improves completion rates. When no suitable tools are available, the model must still possess strong reasoning and planning abilities to find alternative paths.

Co‑training is about making the model better adapt to real‑world tasks. We perform post‑training around specific scenarios, aligning training and online execution environments through an integrated train‑and‑inference workflow. Some tasks perform poorly on the first try, but a slight adjustment of the pipeline can meet expectations, indicating that the bottleneck is not necessarily the model’s capability—it may also be the planning and tool invocation methods. Training and system design need to jointly bridge this gap.

03 From a technical architecture perspective, what core capabilities does a truly useful Personal Agent need? Compared with the relatively mature Coding Agent, where are the technical challenges of a Personal Agent?

The difficulty of a Personal Agent lies in the fact that user context and service ecosystems are more fragmented than those of a Coding Agent. A Coding Agent has a codebase as a relatively complete context, and tools such as editing, compiling, and running are also more centralized, making it easier to form a complete execution chain.

Everyday life and learning scenarios don’t have such natural conditions. For example, when planning a trip, booking tickets and hotels, users often have many implicit default requirements. Even if an agent can execute the task, it may deliver poorly because it doesn’t understand the real needs.

Therefore we believe that a Personal Agent has three core components: model, memory, and ecosystem. The model handles understanding, reasoning, and planning; memory helps it maintain long‑term knowledge of the user; the ecosystem provides expertise, tools, and services. Missing any of the three will affect delivery.

When the ecosystem is not connected, you can’t rely on stronger models to automatically solve it. Direct access via API or MCP is best; if using other methods, it depends on whether the user authorizes and whether the service provider permits it. These integration boundaries are issues that a universal Personal Agent must confront.

04 Many Agent products are impressive at launch, but after long‑term use they reveal unstable task success rates. What solutions does Qianwen have for this?

To impress users, it relies on the upper bound of capability; to keep users using it long‑term, it relies on a stable lower bound of capability. When users use Qianwen to book tickets, they assume most tickets can be booked and will not accept being unable to complete the task if they switch to a different ticket. Therefore, we need to conduct post‑training around specific scenarios to improve stability and coverage.

For example, first understand the demand distribution, then break the demands into trainable atomic capabilities. Although the same type of demand may be expressed differently, they can be clustered and reduced technically. Qianwen currently focuses on more than 30 industries, defining over 8,000 model atomic capability requirements, and improves stability and performance through post‑training and reinforcement learning.

These fundamental capabilities, when combined, aim to cover real user needs as comprehensively as possible. For an application, the ability to reliably complete common tasks is sometimes more important than showcasing the model’s theoretical limits.

05 Personal Agent requires repeated reasoning, tool invocation, and handling an ever-growing personal context. As users engage more, will costs keep rising? What engineering optimizations has Qianwen achieved?

The cost is indeed an issue: when the Agent helps a user snag a ¥10 coupon, the token cost incurred may actually be higher. In the long run, we expect token costs to keep decreasing, but the product itself must also be continuously optimized.

In addition to model routing, model size switching, and inference optimization, the harness is also critical. For the same task, using different harnesses can lead to vastly different execution paths and costs. Whether tools and Skills are laid out flat or layered, and which similar tool is prioritized, will affect efficiency.

We will work with Skill providers to clarify functionalities and applicable scopes, improve the success rate of single-step calls, and reduce retries and context accumulation; we will also encapsulate common execution patterns into Skills to avoid repeated model inference.

Another key point is context compression: historical dialogues are distilled into memory, short‑term Context is filtered by task, and tool and Skill information is also streamlined. Through these optimizations, over the past period, the context size of Qianwen‑related tasks has been reduced by about 40%.

06 In the real world, tasks often have no fixed process. For example, when customer service cannot be reached or order rules change, the agent must find a new solution path. How does Qianwen improve delivery stability for long, multi-step tasks?

Whether long tasks can be delivered reliably depends on keeping the goal, having reliable steps, and being able to change course after a failure. Qianwen Go! Here’s a flight‑delay compensation case: the user wanted cash compensation, but the airline initially offered vouchers. The agent didn’t treat this as a completed task; instead, they kept communicating around the cash goal, looking for other solutions, and ultimately helped the user recover 70 yuan in cash.

This case illustrates the three key aspects we focus on: whether the plan is reasonable, whether each step can be executed successfully, and whether we can replan after a failure. The execution path can be adjusted, but it must not deviate from the user’s goal.

The stronger the execution capability, the clearer the operational boundaries must be. For actions that have real impact, such as orders, transactions, or personal data modifications, user confirmation is required whenever confirmation is needed. We describe interface behavior and interaction conditions in the Skill, and then the system framework imposes strict constraints, creating a double safety net. An Agent acts as an assistant and cannot fully replace the user in exercising rights or assuming obligations.

07 Personal Agent needs to understand a person over the long term, but users generate a massive amount of information every day. What exactly should be remembered, and at what granularity should it be stored? What different approaches does Qianwen have in designing the Memory Schema?

The core of memory design is not to store as much as possible, but to decide “what to remember and how to use it.” Storing too much leads to over‑memorization, while storing too little prevents understanding the user. Even if the schema is not explicitly presented, the content, granularity, and hierarchy still need clear criteria for judgment.

We define facts and relationships, and also infer which memories are valuable from user needs. For example, a user has talked about retirement planning and also about their thirty-something child being unemployed. Each of these is a fact on its own, but together they may indicate that the user is experiencing family-related anxiety. Later, when they discuss family again, or even say “I’m in a bad mood today,” this background can help the Agent understand the issue more accurately.

In terms of storage format, plain text remains viable because the model itself has strong text comprehension abilities. What deserves more investment is content selection, granularity, hierarchy, and timeliness: which information is core, which is only valid at certain stages, and how to update it when circumstances change.

08 Storing memories is only the first step. Some agents remember a lot, but they frequently bring it up in unrelated conversations, making users feel disturbed. How does Qianwen determine when to retrieve memories and which information to retrieve?

Recalling memories isn’t better the more you do; you need to judge whether they are relevant to the current task. Too much information introduces noise, causing the Agent to bring up old matters at inappropriate times; too little use, and users won’t perceive the value of memory.

For example, if the previous user asks about a child’s employment, pension information may be an important background; if they are just asking about the weather, there’s no need to mention family circumstances. The model needs to understand the relationships between events, determine which historical information helps solve the current problem, and how to use it.

Therefore, we will design a memory system based on event associations and causal relationships. This can be understood as a relational structure, but it does not necessarily mean using a knowledge graph; ultimately it may still be stored as plain text.

Memory must also be updatable and expire. Users’ situations and preferences change, so the Agent cannot always rely on outdated information to understand the user, and must also determine whether recalled information is still applicable.

09 In an ideal personal assistant, not only can it execute tasks, but what’s more exciting is its ability to proactively discover needs and provide proactive service. However, proactive service may also lead to misjudgments and privacy issues. To what extent has Qianwen achieved this so far? How can we define the boundaries of proactive service?

Proactive services are still in a relatively early exploratory stage. We first choose scenarios with clear needs and timelines, using product design to limit the scope and avoid excessive intrusion. For example, rehabilitation or follow‑up checkpoints in health issues are well suited for periodic reminders.

Reminders need to answer both “when to remind” and “what to remind”. This requires professional knowledge, an understanding of event relationships, and, combined with product boundaries, keeping reminders within a reasonably appropriate range.

Related does not mean it should be repeatedly brought up. If a user has consulted about a disease they do not wish to discuss much, even if it is related to subsequent health issues, it should not be reminded each time. We will use text and semantic constraints to clarify which services can be provided and which reminders should not be generated.

The clearer the long‑term goal, the easier it is to follow up. When a user sets a marathon target time, subsequent training, running, and rest can all be linked to that goal; but “I want a promotion” is a very broad scope, and it’s not yet possible to reliably judge how current tasks serve that goal. At the same time, the assistant must respect the user’s immediate needs and cannot intervene with the long‑term goal every time.

So at this stage, we first focus on clearly defined tasks and goals, providing appropriate suggestions and necessary information along the way. By consistently achieving these, we can gradually move toward more complex proactive services and long‑term goal management.

10 What are your expectations for the future development of Personal Agent?

What I look forward to most is that the Personal Agent not only executes existing tasks, but also creates new solutions: discovering paths the user hasn’t considered, or helping the user accomplish things that would otherwise require a lot of time and cost. I believe this is its true incremental value.

Today, coding and work‑related agents are developing rapidly because the inputs, outputs, and context are relatively clear. Broader opportunities may lie in everyday scenarios, such as buying a house, buying a car, or children’s schooling. These problems have fuzzy boundaries yet consume a lot of time and energy, and they are a key focus of Qianwen’s exploration. By combining personal circumstances, professional knowledge, and external services, agents can potentially provide better solutions.

We are moving from agents that complete well-defined tasks to assistants that create new value, but this process is still in its early stages. I hope to first see some truly useful everyday scenarios, and then further proactively help users get things done. To put it bluntly, either help users earn money or help them save money.

© THE END 

For reprints, please contact this public account for permission

Submit articles or request coverage: liyazhou@jiqizhixin.com