On October 31, the DeepSeek Harness team released a long-form, super-agent-designed search engine that quickly brushed through the AI circle.
The article discusses a dilemma that many people have experienced:The more agents are used, the more busy people are.
According to DeepSeek, a person can usually only follow three to five Agent chat rooms at a time.
The upper limit of a single task depends on the ability of the model and the upper limit of a multi-task depends on the attention of the person.
They envisioned a long-term super-agent online, actively interacting, assigning tasks to other agents and returning to a decision-making and acceptance position.
Dual-layer architecture: Router layer unified routing with diplomatic interaction, multiple Worker Session isolation and simultaneous execution of tasks
There is also an inspiring line in the text:
Similar design constraints lead to similar architectural choices.
With that, we scrambled through similar products that were already online and finally found the Chinese open-source company LobeHub.
The company, which has garnered more than 84,000 stars on GitHub, has spent most of the past year experimenting and organizing its own research and development using its own Agent platform.
According to its founder, Manchu, the team has not launched the development server locally for three months, and after seeing the acceptance report submitted by Agent, it merged directly, increasing the development speed by 3 to 5 times.
Instead of translating these statements, we wanted to verify them in person, so we arranged a night shift for LobeHub to have Claude and ChatGPT make up a two-person editorial department and give us a customized AI diary every morning.
The following is a complete transcript of what happened to the LobeHub team.
Two AI editors were sent live the night before.It's all right.
We have three members of the editorial staff at LobeHub: Agent in charge of posting and receiving, Claude in charge of collecting material, listing outlines and writing correctly, and ChatGPT as reviewer in charge of pointing out flaws in logic and perspective, and finally creating covers and diagrams for articles.
The reader's portrait is written by us, a tech editor who focuses on topics such as Agent, Key Models, AI, Commercialization Progress, and spends five to eight minutes a day reading the top five to eight stories of the day, each with a source and a sentence or two.
In the process, we intentionally leave the judgment to Master Claude.
The GPT may issue an opinion or Claude may refuse, and as long as the reasons are given, the parties come back to the discussion until the GPT expresses no significant objection, or until all four rounds are completed.
In general, our design thinking is:
- Claude is responsible for collecting material and drafting the outline - > ChatGPT
- Claude refers to the fact that, in his own discretion, he decides whether to modify the outline, collect new material, form opinions, and then submit them to ChatGPT, which debates whether to merge into a unified opinion.
- Claude writes a preliminary draft
- It is not necessary to debate whether or not to modify the font this time, as the readability of the text is much stronger than the GPT, but this piece of previous writing thought can help to supplement the lack of perspective and logic.
- After completing the modification, Claude handed the article over to ChatGPT, where the correspondingly illustrated section (including the cover) called for the strongest graphic model GPT-Image-2.5-Sunburst (because before the screenshot LobeHub was not yet online supporting the GPT-Image-2.6 that was just released yesterday).
- Delivery of documents
LobeHub's own team has been working on this multi-agent collaboration for the better part of a year.
Their research and development team is headed by Agent Nasa, who is responsible for viewing mission tables, allocating work and reporting progress.
Early execution agents could speak directly in the group, several agents reported in turn, and the group was quickly filled with noise, and later the team set the rules, with the execution agents reporting only to the supervisor, unified by the supervisor.
Now they've added another layer, each with a Personal Agent who filters information for themselves, and they only listen to their Personal Agent, who is sent by the project manager.
It's not always the case that the Executive Agent is someone else's mouth.
At 8:24 p.m. on October 23, the founder of the group saw a screenshot of a detailed training page, commenting that "the UI looks too vibe at first glance to think about how to refine it", and his colleagues went to dinner and no one followed for 55 minutes.
At 9:19 a.m., a design agent decided that this was related to himself while searching the channel, taking over without being labeled, pointing out that the progress information on the page was repeated three times, the frames were framed, and the most prominent position was empty, and at 11:52 p.m. he issued a contrasting draft of the changes before and after the changes.
DeepSeek envisions a proactive response, filtering the rules first, and then judging whether the model intervenes, and the idea of both is very close.
In the early hours of the morning, two editors debated the issue.It's all right.
After the master agent assigned the task, we turned off the computer to sleep, and Claude and ChatGPT continued the discussion in the cloud.
AI and AI are reviewing each other in the middle of the night, and it's already the norm in the code warehouse at LobeHub.
At 4:04 a.m. on October 31, Executive Agent Dario responded to the comments of the OpenAI Codex review robot under PR #21852.
The code came from NASA's T-1149 mission, which took Dario 31 steps and took 1 hour and 56 minutes to complete.
For the next 18 minutes, it corrected all four issues and justified another opinion, deciding to keep the original writing.
These efforts can continue in the middle of the night, based on the architecture of LobeHub, which went online in January this year.
Early product deployments on Vercel took only 300 seconds to execute, and the team divided each step of the Agent into a separate function call, using Redis to preserve the middle state, power up thousands of long tasks, and send the tool call to the AWS sandbox or the user's own computer to execute.
Agents do not perceive their own thoughts and actions taking place in different places, which is the same direction as the "computing, state and environmental separation" in the DeepSeek text, which today stably carries tens of thousands of Agent executions per day.
In the early morning news, people only see the results.It's all right.
When we got up in the morning, we got the results of the first edition of the daily newspaper for my review.
In this first edition of the newspaper, what did you find wrong?
Some of the text's conclusions seem to be quite rigorous, but the plot is unclear, so the standard is "beautiful garbage".
As the boss of the Agent, I quickly took a breath and gave a suggestion for modification:
If it doesn't have a particularly worthy real picture (such as the new model's runway versus the chart), then you can use foreign news to artistically process the image of its founder (the soul character) to represent the image of its founder: this new news can be very healthy if it's a big benefit to the company, it can be exaggerated, detached from reality, as long as it can show how big the progress is), it can turn into a muscular rapper; it has no impact on the original image; it can be depicted as a smaller picture than a bigger picture, and it can be depicted as disappointing, but it can be depicted as a smaller picture than a bigger picture.
Nothing else needs to be changed.
My agent's manager said it was his mom, and he changed.
We human users just look at the results and score points to find out exactly how we made such a stupid mistake, and we just have to say yes or no.
People just look at the results, and that's how the LobeHub team works now.
On October 29, after the release of v2.2.21 version merger, Agent continued to follow up on the construction of the platforms, and at 8:26 p.m. discovered that the desktop stable version construction failed, and the reboot was not yet approved, and was located in a dependent version conflict, directly opening the repairable PR #21910, just to ask for a confirmation.
Acceptance is an open-source acceptance tool from LobeHub that saves people time.
Before the agent is delivered, the requirements are broken down into inspections that can be observed directly, run one at a time in the real product, combined with screenshots, recording screens and logs, the person sees the report decides whether to pass or return, and the part that has been passed does not need to be returned.
Code testing can prove that a program can work, but there are many details that need to be used once and for all, just like a real user.
The previous PR #21852, and the last to wait for only one human operation, pressed the merger button at noon on 1 November, and the task was completed.
The following day, the newspaper recalled yesterday's comments.It's all right.
GPT teacher: You can't really sit down, that's the mystery of the thunder muscles you want.
Sam Altman after the release of GPT-6 Cyber (imagined by AI)
Agent finally remembered what kind of AI news we wanted to see, and we were delighted.
From then on, at 8 a.m. every morning, and no more folding like the first time, LobeHub would send me our favorite AI diary via mobile phone, so we could choose a topic on the subway.
Remembering users' opinions is just the first step in the RSI.
RSI, or recursive self-improvement, which refers to systems continuously improving themselves using the results of their operations, is one of the most discussed concepts in recent AI circles.
This test first confirmed one thing: Agents can retain user feedback and use it to adjust the rules for follow-up work.
But for the LobeHub team, this is just a small snippet of their exploration of continuous improvements to the system.
In internal research and development, they have attempted to automatically collect, classify, and convert issues exposed during Agent execution into new improvement tasks;
Allow the Agent to autonomously optimize the Harness when running and verify the results through benchmarking tests;
It also explores how to use real execution trajectories for vertical modeling.
The work takes place at the three levels of tasks, operating frameworks and models respectively.
LobeHub hopes to connect these links gradually so that the system can continue to improve itself using feedback from real-world operations.
Next stop, L4.Gartner predicts that by 2028, 33 percent of enterprise software will have Agentic AI built in, compared with less than 1 percent in 2024.
When an agent is transformed from an individual tool to an organizational member, what is really scarce is the layer of collaboration that allows many agents to work together.
LobeHub compares this route to self-driving:
- L1 is a question-and-answer tool.
- L2 is a personal assistant.
- L3 is a collaboration between people and multiple agents on the same site.
- L4 is an autonomous interface between agents, where people only set goals and boundaries.
- L5 is an organization's self-evolution in operation.
They believe that their team is in L3 and that the upstream project's Agent directly addresses the needs of the component repository lobe-ui, adding the component automatically and then notifying the upgrader, which is already the default of L4.
The bottom line behind this vision is that it has accumulated a little bit in three years.
LobeHub started with LobeChat, which was open source in 2023, and Docker has been deployed over 7 million times.
DeepSeek also used LobeChat to demonstrate its model capabilities when it released the explosive R1. Open source allows it to have an existing distribution channel, model neutrality allows all models to have access, and tens of thousands of times a day to execute submerged data.
DeepSeek said in its Super-agent technology report:Multi-tasking limits depend on people's attention.
Our daily newspaper goes from writing to writing, and all people have to do is say what they want the night before and judge it badly the next morning.
We reviewed the operational records.
We've been forced to do it ourselves a few times before and after, and most of the time we've spent the authorization on the first configuration, the quota isn't enough to replenish the quota. AI: It's product design, sir, and you don't want your bank card blown up by me, right?
From the actual experience, the LobeHub team is in a state of merging only with a single button, and there are still several gates between ordinary users, and there is still room for hands-on experience.
But, well, saving the energy of switching, converting and comparing between multiple windows would only make it more frustrating.
The precious energy of this human user has not disappeared, but has been diverted to the exchange of needs and judgments of the results, and is used on the blade.
Today's Agent has taken over the middle execution, start-point direction and end-point swap, and it's still up to people to come up with ideas, and what LobeHub wants to do is make the path between the two shorter and shorter and, together with users, smooth out these boring thresholds.
