The afternoon before last, the world's first enterprise-level AI Agent with 'memory', Zhipu CoCo, quietly began its closed beta.
CoCo is billed as an Agent that 'understands you and understands enterprises, can do work and can deliver'. It is not just an Agent that helps you 'write something' or 'look something up',but an Agent that can fully handle complex tasks and deliver professional results that can be used directly。
In this article, we will take a deep dive into CoCo's core capabilities together. Through test cases from three real industry scenarios, we will analyze how CoCo can bring efficiency leaps to enterprises.
01. General-purpose functions within enterprises
First, let's look at policy interpretation.
Prompt: The State Council recently approved the 'Work Plan for Accelerating the Comprehensive Pilot Program for Expanding the Opening Up of the Service Industry'. Please analyze what impact this will have on China's tourism industry, and help me generate an illustrated interpretation.
Unlike most agents, CoCo does not directly complete the task,but instead sorts out a clear and specific execution plan, waiting for me to edit and confirm before proceeding with the task。
The plan CoCo provides is detailed and comprehensive. It lists as many as 10 analysis dimensions alone, and is completely ready to be adopted as is.
CoCo only needs to help us sort out a clear and specific execution plan, while our action of clicking the confirm button is much more taxing.
During task execution, we cansee the task progress bar in real time, as well as CoCo's detailed work log。
For ordinary agents, when collecting information, even if the prompt requires confirmation of information sources, hallucinations still easily occur. The collected content often needs to be manually reviewed and confirmed one by one, wasting a lot of time.
CoCo's standard for information collection is:prioritize retrieving 'authoritative', 'authentic' 'original' content. It has very high requirements for information sources, so naturally the quality of the information collected is also very high。
After the planning node is completed, we can directly view its data sources. They are all relatively well-known official media. Great review!
After all plans are executed, CoCo delivers several files, each of which can be clicked open and viewed separately.
Let's directly open the final generated webpage and take a look:
The overall layout of the page is very simple, and the data presentation is very intuitive.
The content of the webpage is based on the interpretation of the policy, as well as the analysis report made by CoCo. The content is very rich.
For some key data in the webpage, the data sources are attached below, and clicking will directly jump to the information source link.They are all relatively authoritative official websites or media.
It can be said to be 'well-founded and convincing'.
02. Role-specific functions for various positions
Let's try again and have CoCo complete some work for specific positions, such as helping me recruit an assistant editor.
Prompt: I am a tech self-media blogger focusing on the AI field, sharing AI practical content and use cases, and committed to helping more people turn AI into their own productivity. I hope you can help me recruit an assistant editor to assist with the content output of the official account. City is unrestricted, education is unrestricted, and some understanding of and interest in AI is required.
CoCo also directly lists a plan.
This time I want CoCo to do a bit more, preferably directly putting suitable resumes in my hands. So I added one item to the plan:
Finally, CoCo generated the core ideas of the recruitment plan, the job description, and the recruitment copy for me.
The content was still quite good. I did not describe the specific job requirements, and CoCo directlybased on similar positions on recruitment platforms and recruitment information from leading self-media bloggers, automatically helped me sort out the job responsibilities and recruitment requirements for this position。
But it did not help me collect resumes, possibly because I did not provide/log in to accounts on major recruitment platforms.
In its reply, CoCo mentioned that if there is a subsequent need to push candidate resumes, it will process them based on the actual resume materials received.
So I directly prepared 10 resumes and asked CoCo to help me screen them.
Prompt: I have received these resumes. Please help me screen out suitable candidates from them.
After receiving the resumes, CoCo will first compare candidates' abilities horizontally across multiple dimensions, including AI interest/experience, content editing/operations capability, AI tool application, written expression, and remote/multi-platform experience.
In the end, it recommended 3 relatively suitable candidates to me.
After looking at these three resumes, they were indeed a better match among the ten resumes, which shows that CoCo did a pretty good job.
Of course, building an excellent team definitely requires talent in all aspects, so I continued to ask CoCo to recommend more.
I also need a finance person and a legal person. Do you have suitable candidates?
This time CoCo's answer was that there was only 1 candidate highly matching the XX position, and the reply was very accurate.
CoCo introduced a unique memory mechanism that can remember each employee's functions, work priorities, etc., and provide different AI services for different employees.
During my evaluation process, CoCo was recording my work style and work content, andthe memory duration can be adjusted, up to permanent。
When we use different users to generate the same content, the results CoCo gives are also generated based on your position and habits in memory.
For example, for two different positions of ours, such as R&D and new media operations, each generates a daily financial report for today:
R&D personnel's financial daily report
Data analyst's financial daily report
We can clearly see that these two daily reports are both very targeted. CoCo adjusted them to better suit each position.
With the support of this unique memory mode, CoCo can continuously understand the context, adapt to you, and feel smarter and more understanding when used.
We can hand over some work with clear requirements and repetitiveness to AI, helping enterprises recruit suitable candidates at the lowest cost and in the least time.
03. Customize personalized tools
What surprised me most is that for tasks that need to be executed repeatedly, CoCo cangenerate a workflow in one sentence, and form a repeatedly callable MCP service.
It is equivalent to being able to customize a "mini program" in one sentence according to your own needs.
For example, in a while, many friends will have the need to fill in college entrance examination preferences. We can ask CoCo to generate a workflow. We input information such as scores, and CoCo recommends colleges and majors for us to apply to.
Prompt: Help me make a workflow. Based on the selected subjects, province, and estimated college entrance examination score, recommend 4 colleges and majors each according to steady, safe, and reach categories, and based on the recommended school and major list, generate illustrated introductions to the schools and majors.
After confirming some details with me, CoCo will generate the workflow with one click.
You don't need to choose nodes yourself throughout the whole process, and there's no need for wiring or layout...
How is it? Pretty convenient, right?
Once a workflow is built, we can call it anytime in chat.
For example, the workflow we just generated, I named it "My College Application". When we need to use it, we just enter in a new chat box:
Prompt: This is my gaokao information: Hubei Province, History, total score 470. Use the "My College Application" tool to recommend schools and majors for me.
CoCo will then, based on the previous requirements, call this "My College Application" tool to recommend schools and majors for me.
With one sentence, CoCo works away busily. We have our own digital employee now, hahaha.
04. Some sharing
CoCo's core value is not to replace certain positions, but tolet everyone unleash greater value and help enterprises reduce costs and improve efficiency。
It can alsocombine with an enterprise's own database or knowledge base, to produce more customized solutions and content.
CoCo supports fully private deployment, calling enterprise local models, connecting to enterprise intranet tools and resources, and running entirely in a local environment, fully meeting enterprises' data security needs.
However, CoCo is currently still in beta and requires an invitation code to use~
Beta address:
https://aiworker.aminer.cn/ai_worker/verification
Get a free Zhipu CoCo invitation code, scan the code to follow and leave a comment in the article comment section
Original link:CoCo: The world's first enterprise-level AI Agent with "memory"