This "Bar-headed Goose" just set a new single-round funding record in the Agent space: Who is it? Where did it come from?

**

We Built a "Tangdao GuaGua Agent" With It

👦🏻 Author: Jingshan

🥷 Editor: Kavana

🧑‍🎨 Design: NCon

The bar-headed goose is the only bird species capable of flying over Mount Everest. While other birds can barely pass through with the aid of wind currents, the bar-headed goose can power its way across the world's highest peak on its own strength.

This more or less reminds us of the current state of the enterprise AI Agent market: most products are waiting for the perfect "wind" — stronger models and compute power, lower costs, a more mature technical environment — but there are also products that refuse to wait for perfection and instead seize the day, attempting to "power through" the peak like the bar-headed goose.

Recently, an enterprise AI agent development platform that named itself "Bar-Headed Goose" (BetterYeah AI) completed a Series B funding round of over 100 million RMB led by Alibaba Cloud, setting the record for the largest publicly disclosed single financing round in this domestic sector to date.

BetterYeah AI's core team has deep ties to Alibaba. CEO & founder Zhang Yi (alias: Taojun) was a founding member of Alibaba's DingTalk and former VP; the other two co-founders (COO Huang Wen and CTO Huang Chongkun) also came from DingTalk's core founding team.

🚥

With AI Agent concepts flying everywhere today, what is the standard for measuring whether an enterprise-grade product has the ability to "power through"?

The answer may lie in just four characters: "deployability."

We found that BetterYeah AI scores quite well on deployability — the hardest hurdle for Agents.

Therefore, we conducted an in-depth review of this commercially notable product, and used three key insights to answer the question: "After the AI Agent hype dies down, what kind of product can actually survive inside enterprises?"

We Put the "Tangdao GuaGua Agent" to Work!

BetterYeah offers AI Agents and workflows for various scenarios that can be used directly as references. We also tried building an AI Agent for Tangdao ourselves — the "Tangdao GuaGua Agent."

1) Robust Workflows

Our vision for the "Tangdao GuaGua Agent" was: an AI customer service agent with industry-specific expertise, capable of calling tools, that delivers customized responses to customers — one that actually gets work done.

Let's see if it can deliver.

First, we can set up a role definition based on what we want it to do, establishing its foundational capabilities. For the role definition section, I created visual cards placed on the right side — feel free to copy them directly:

Then, BetterYeah makes it easy to upload data to knowledge bases and databases. For example, we uploaded three Tangdao products: the Tangdao Line Friends Summer Cool Quilt, the Tangdao Mountain Melon Summer Cool Quilt, and the Tangdao GuaGua Cool Quilt.

After using AI to format their respective information, we can copy it into BetterYeah's knowledge base with one click:

Knowledge base uploads are convenient and support multiple formats

This gives us a basic version of the Agent.

For example, when we ask it:

Prompt: I live in Shanghai, I'm afraid to run the AC overnight in summer, and Shanghai occasionally has plum rain weather. But I absolutely need to sleep under a blanket to feel secure. Everything feels so hot in summer! Please give me a recommendation?

The "Tangdao GuaGua Agent" can draw from the knowledge base data, engage in deep reasoning, and provide a very complete response with four recommended reasons, plus product images and purchase links from the knowledge base — clicking directly takes you to the Taobao interface:

We won't dwell on BetterYeah's most basic features.

Beyond that, in our actual use we discovered:

This product's biggest advantage is its ability to handle highly customizable, rich scenarios, because it supports building very personalized workflows and assembling these flows into Agents that can quickly be put to work.

2) Highly Customizable, Rich Scenarios

Weather Queries

When users face scenarios where customers might need "weather queries," they can quickly build a weather query workflow on their own.

BetterYeah provides a canvas-style workflow building interface, and at every node in this workflow that requires AI capabilities, you can independently select which foundational large model you want to use.

BetterYeah offers extensive model options, such as DeepSeek-R1, V3, Tongyi's Qwen, and more — each with different point consumption rates.

The left panel provides numerous foundational capability nodes and personalized templates, including tools for Bilibili, Weibo, Douyin, and more — this is actually quite similar to the "MCP" concept:

After building the workflow, we can give it a "description" to make the AI capabilities more precise:

We input a prompt to test:

Prompt: What material quilt would be suitable for Beijing's current weather? Any recommendations among Tangdao's summer cool quilts?

The "Tangdao GuaGua Agent" will then automatically call the "weather query workflow" based on its reasoning, retrieve Beijing's temperature, and provide a recommendation accordingly:

Refund-Only User Intent Recognition

Since BetterYeah's workflows provide LLM nodes that can directly leverage large model capabilities for content parsing, we can simulate a tricky niche scenario: recognizing the intent of users requesting refunds without returns.

The approach to handling this scenario is actually quite simple:

For example, in this "refund-only" workflow, we divided user intents into four categories:

Then using LLM nodes, we set a separate prompt for each intent category.

For "users with active claim intent," the LLM needs to consider shipping insurance information comprehensively and output a response template to guide the Agent's eventual reply to the user:

Prompt: If the user is "Active (claim intent)", output: 💬 "Hi there, GuaGua understands your concern 🥺, but according to platform and store rules, refunds require receiving the returned item first~ If shipping costs are inconvenient for you, GuaGua can help check whether you qualify for shipping insurance or see if other subsidy options are available. How does that sound?"

Then we name this "refund-only workflow": "Refund-Only User Response Strategy Expert," with the instruction "when a user wants a refund without return, this skill should be called":

Then we simulate a test case:

Prompt: I have to pay for shipping myself? Then just refund me the money.

"Tangdao GuaGua" will then use the "Refund-Only User Response Strategy Expert" workflow to identify the user intent as "active claim type, focus on shipping costs and refund":

The user is clearly dissatisfied with the shipping cost issue and may be emotionally agitated. I need to address the user's emotions first while following company return policies. According to the execution results from the Refund-Only User Response Strategy Expert workflow, the system has already provided a response template emphasizing empathy, policy restatement, alternative solutions, and transferring initiative. I need to ensure my response covers these points.

The Agent first gets its direction from the auto-generated response template in the workflow, but doesn't copy it verbatim — instead, it builds on this foundation to provide a more expansive, contextually appropriate answer:

AI Image Recognition

We can continue DIY-ing by using the "AI image recognition" function in BetterYeah's workflow nodes to build an "image upload analysis workflow" that receives user messages and uploaded image files, analyzing and processing the image content.

In the "Start" node, we first have the user input text + image, then process it through "video/image Q&A" analysis, and finally search the knowledge base for related products:

We upload a random summer cool quilt image from Taobao and ask "Tangdao GuaGua":

Prompt: Do you have any recommendations in a similar color?

The "Tangdao GuaGua" Agent identifies in its reasoning:

The Mountain Melon Summer Cool Quilt in the knowledge base is a green color scheme with a natural mountain wilderness pattern, which may match the color in the user's image. Need to check whether the product description's color and pattern are similar to the user's uploaded image. Additionally, the user likely prefers a natural, fresh style, so the recommendation should emphasize color and design elements.

Finally, it gives the most suitable answer — the "Tangdao Mountain Melon Summer Cool Quilt":

3) High Extensibility

Of course, we believe many who have actually used it have also discovered BetterYeah's "hidden use case": because BetterYeah AI has very high extensibility, Agents built on this platform can essentially be connected to and used anywhere.

From a usage scenario perspective, beyond common web links or direct chat conversations, users can also publish built Agents to various different platforms for use.

BetterYeah also provides MCP services, supporting turning completed Agents directly into MCPs and connecting them with one click to external AI Coding IDEs, using BetterYeah's own operational capabilities to do "multi-Agent nesting."

To give a simple example, we can manually configure a "Tangdao MCP" composed of the "Tangdao GuaGua Agent" in the IDE.

BetterYeah directly provides the URL for one-click configuration:

Then, in the IDE, build an "upgraded version of the Tangdao GuaGua Agent."

For example, directly give it a prompt and add "Tangdao MCP" in the tools:

Prompt: You are a Tangdao GuaGua Agent. When the user inputs "please recommend a summer cool quilt for me," you need to use "Tangdao MCP" and output the results. Automatically generate an html webpage from the search results.

Then this ultra-simple version of the "Tangdao GuaGua Agent" can automatically output visualized web pages:

This approach of turning BetterYeah Agents into MCPs allows seamless integration into professional development environments.

At this point, developers can add their desired interfaces, features, and interaction methods on top of mature BetterYeah AI Agents.

4) Tangdao Custom Summer Cool Quilt Shipping System

BetterYeah actually also has a multi-Agent collaboration feature that allows users to chain together more complex scenarios.

To explain this function intuitively, we tried simulating a "Tangdao custom summer cool quilt shipping system." This task in real-world scenarios typically requires a very long process — for example, there's a user "Mr. Zhang" who customized a Tangdao summer cool quilt with his own preferences.

The process involves the following steps:

[1] Order query → find Mr. Zhang's order number
[2] User preference parsing → interpret "back sleeper, prone to neck pain, allergy-prone" as specific materials
[3] Factory capability matching → check whether such materials (e.g., anti-mite fabric, slow-rebound fill) are in stock
[4] Shipping confirmation → review custom order completeness, send to production and arrange shipment

However, each step actually involves multiple workflows for the Agent, so at this point we can build several sub-Agents to form a main Agent, breaking down this complex scenario.

We can quickly build an Agent for each step using BetterYeah.

For example, upload order data and build an order query Agent:

Order data

User Preference Parsing Agent:

Factory Capability Matching Agent:

Once the above three sub-Agents are built, you can build a main Agent in the lower left.

After checking all three Agents with one click, the "Tangdao Summer Cool Quilt Customization Agent" is quickly complete. We can input a prompt to test:

Prompt: Can Mr. Zhang's custom summer cool quilt be arranged for shipment? Give me the order number, preference data, and material readiness information.

You can see that the main Agent will query the 3 sub-Agents separately, obtaining information in sequence, then arrive at an integrated answer through AI's deep reasoning capabilities:

Finally, every built Agent can have full-process data monitoring through logs, analytics, and performance monitoring:

And this is just the simplest case of BetterYeah's Agent collection feature — many more possibilities are worth exploring.

The Hardest Hurdle for Agents — Deployability

Currently, various excellent AI Agent products are springing up like mushrooms, from chatbots to AI customer service, from AI Coding IDEs to writing tools, in a process of "letting a hundred flowers bloom." But close observation reveals that products that can truly take root in enterprises, be used in daily operations, and create productivity are somewhat "few and far between."

The hardest hurdle AI Agents face in achieving true commercial deployment is: can they actually "deploy," can they solve real problems.

It shouldn't just be a technical prototype, but a real commercial product.

1. In the Agent's "Green Zone" Rather Than "Red Zone"

Some time ago, in a paper on AI Agents from Stanford University, I noticed some quite interesting insights: AI Agents should focus on work in the "green zone," not the red zone.

What is the "green zone"?

Simply put, it's those task domains that both need automation and are easy to automate. Examples include business negotiations and financial analysis, data processing and computational work, and various management and coordination affairs.

This isn't just because AI can replace human labor and create profit — more importantly, the "colleagues" working alongside the Agent can accept this collaboration method.

The "red zone" is where AI can indeed achieve high automation, but user mental demand is lower — examples mentioned in the paper include logistics analysis, psychological counseling, courts, and other domains.

In the Agent's "green zone," enterprise AI Agents can maximize their effectiveness. They possess professional industry knowledge, allowing expert-level productivity to replace repetitive workflows.

In scenarios requiring professional industry "context," task confirmations with different permission levels, and complex collaboration, enterprise AI Agents can find their proper place, solve real pain points, and significantly boost productivity.

2. Embed AI Capabilities at the Right Position to Amplify, Not Replace

If we closely observe the AI Agent products that have achieved initial success, we discover a common characteristic: they don't simply aim to "replace" users, but smartly choose to "enhance" user capabilities at the right nodes.

We've seen many AI products demonstrate in their interaction design: the thinking of "cleverly embedding AI" when building workflows.

Taking BetterYeah AI's workflow construction as an example:

[1] Inserting retrieval-enhanced, parsing-enhanced AI modules at certain points in the workflow

[2] The "role persona" in Agent capability settings

[3] Description words in workflow call configurations, and so on.

Multi-layered AI capability nesting actually also means that in some places, we may still need to admit: AI isn't that smart yet.

For example, AI performs well in intent recognition for "refund-only" users; but in single-turn conversations with limited context, such as inquiries about discount information, a simple line of JavaScript or just directly searching the "coupon knowledge base" for relevant information may be more effective:

The user decides the goal, controls the pace, and makes the final judgment, while AI simply helps us do each link better, faster, and more accurately.

3. Fully Leverage the "Abundance" and "Accessibility" of Foundational Large Models

Thanks to substantial price cuts by foundational model providers, we believe enterprise Agent products are also entering their moment in the spotlight — this cost reduction has brought unprecedented opportunities for enterprise AI Agents.

Today's AI Agent products, with user collaboration, are becoming increasingly "smart" through rapid workflow construction: they can flexibly select different "tiers" of AI large models to handle tasks of varying complexity.

This approach has a win-win benefit: users can control costs based on the task itself, while AI Agent product providers can also more efficiently deliver compute resources.

Taking BetterYeah AI as an example, when I did the math carefully during in-depth testing: a team plan user's 20,000 point allowance, if using Tongyi's open-source lightweight models for daily simple inquiries — such as the cheapest Qwen Turbo (0.02 points/call), or choosing cost-effective domestic models like ByteDance's Doubao or Baidu's ERNIE Bot — in scenarios building relatively simple AI Agents, can support 500,000 to 700,000 conversational interactions.

This massive improvement in cost-effectiveness is precisely the key factor enabling AI Agents to truly take root in enterprises and stand out in capital markets.

🚥

In 2024, nearly every AI industry entrepreneur asked the same question: Should we build an AI Agent?

In 2025, that question became: Why hasn't our AI Agent made money yet?

At every technical inflection point in AI Agent development, there are always moments of sudden realization:

The real breakthrough doesn't lie in how "shocking" the technology itself is, but in whether it has found the right commercial soil.

From the BetterYeah AI case, we see a relatively clear path:

[1] Don't try to build a universal AI Agent, but rather a professionally matched "digital colleague";

[2] Don't aim to disrupt everything, but first solve specific problems;

[3] Don't pursue technical perfection, but pursue commercial value realization.

This may be exactly the answer that entrepreneurs still anxious about AI Agent commercialization most need.