From the Temple of Penang
QbitAI is public.
An AI editor leads the three AI writers who work on the front lines of big models, open source projects, and factory staff.
I put the demands on the subjects in the hands of the editor-in-chief, Agent Sienna, and he turned around and dismantled the living, sending people, collecting manuscripts, and summing up the whole story.
Not at all.Eight minutes.It just came to me.Article 6The newly hot-rolled AI selection:
No, my tea hasn't boiled yet, and the day's assignment has been so boring!
It was six months ago.The LobeHub teamThey started doing research and development on their own AI collaboration platform, allowing Agents to participate in assignments, write code, and even scan Bugs.
Now, AI is a permanent member of the LobeHub team's R&D process.
Of FrontierHarness's 30 software engineering missions, 21 were completed using its optimized Agent Harness using the Kimi K3, with a rate of up to70%The average cost of each successful mission isIt's about $1.68 (about RMB 11.21)It's more than the Codex and the Claude Code.
It's hard to believe that DeepSeek has recently been exploring the Super-Agent architecture, and LobeHub has already made a similar idea of human collaboration into a product as early as May, even directly open source!
Today, the LobeHub team has incorporated this approach into everyday research and development, allowing Agents to dismantle tasks, activate, write code, and even help review modifications.
On GitHub, LobeHub is already a craze.844 thousand stars。
So how does LobeHub move Agents from single-handed combat to team collaboration, how is it tacit and to what extent can it move the work forward?
AI develops AI, LobeHub develops itself
AI gives AI life, AI gives AI code.
Such a scenario has become a daily occurrence for the LobeHub R&D team.
It was a development assignment submitted on October 28.
At the time, the team wanted to add a new feature to the task tool and command-line interface so that Agents could directly specify which model to use when creating tasks.
So, they handed it over to the Director Agent.Nasa。
Upon receiving the demand, Nasa immediately created a mission numbered T-1149 and directly sent it to the corresponding coding agent.Dario。
After receiving the assignment, Dario nodded his head to the code.
From implementing the functionality, completing the test, to checking the results of the operation, it successfully delivers a correct pull request, numbered#21852。
△NASA sends live recordings, Dario submits a press releaseBut the mission is not over.
As soon as the code was delivered, another AI colleague rushed in and punched him in the face.
This year's event was held at OpenAI.Codex code scans robots。
When the Codex examined the code, it quickly identified several issues that needed to be addressed.
After receiving the feedback, Dario turned to modify the code, complete the test, and finally return to the task.
In this way, the Codex and Dario live like real-life development engineers and code reviewers, one responsible for writing, one responsible for bugs, and you come to me, and I've been working on the code for a few rounds.
Eventually, the PR was successfully reviewed and integrated into LobeHub's code library.This is how a real AI R&D collaboration chain runs.
△Codex review comments, Dario changes to records, state of PR mergerWhat's even more interesting is that LobeHub's agents aren't even satisfied with single-handed combat.Starting a proactive cross-project search for other AI colleagues.
For example, when an Agent-foundry project requires a lobe-ui component library to add custom neutral color support.
In the past, there was definitely no shortage of developers needing to communicate, coordinate modifications, and then notify each other of updates.
But this time,Agents on both projects are directly connected to their heads.
One submitted a request, the other completed a modification. With the release of 6.0.1 by lobe-ui, upstream projects also received an upgrade notification, so the new features were seamlessly added.
△Interproject Agent communication records, component release recordsIt's amazing how humans used to have to go back and forth between the two projects, and now Agent has turned it upside down.
Moreover, the LobeHub team has really tasted the doughnuts.
According to a research and development member of the LobeHub team, he has not launched Dev Server locally for three months.
According to his personal feelings, the development efficiency has been improved by adopting this approach to work.Three to five times。
It's not uncommon for people to create their own AI that helps them develop their own products, and even start coordinating with other AI colleagues.
LobeHub is actually using this way of working, and at the same time saving a lot of spamming.
Then the question arises.
What happens when I hand over the same set of systems to myself as a qubit programmer?
Wow, I've actually set up an AI editing department with LobeHub!
In order to verify the reality of the AI team's work, and to change the work landscape, I decided to start working with the tech editors that I was most familiar with, and I personally drew an AI editor team and gave them two rounds of action:
The first round.Let the AI Editor-in-Chief, along with the writers, find a selection of topics and produce a complete summary;
The second round.Let's see if the results that we've just delivered can be passed directly to the next player to continue working on, generating a dynamic hotspot web page.
Come on, get to work!
Test One: Let AI be the qubit editor at once
First, I set up a team of miniature quantum bit editors in the LobeHub.
Subtitles by SiennaThere are also three AI writers who are in charge of the governing body.
Gemini 4 ArgonHe is responsible for managing the shift in staff at the plant, and to see who else has jumped in the trenches recently.
ClaudeEmbodying Open Source Star Search, dedicated to excavating new projects in the GitHub community;
DeepSeekIt is responsible for frontline surfing, looking for the latest model dynamics worth reporting on.
Everything is ready and ready to go!
I ordered Sienna to identify two of the latest trends that are worthy of attention for me in the three aspects of personnel change, new model releases and open source new projects;
I'm not sure what I'm saying.
Our readers are a group of 5G surfers who are constantly focused on AI models, leading edge technology dynamics such as physical intelligence.
It's common for people to spend it every day.5 to 10 minutesThe fast browsing industry is moving, so the selections are new, accurate, and information-intensive, allowing readers to quickly see what is happening, why it is worthwhile to be concerned, and avoid basic cop, widespread and pure propaganda.
Please select two dynamics worth a qubit report in each direction, for a total of six.Each includes a candidate's title, news date, core news facts, reasons for recommendation, and a direct open source link.
The first time the mission went out, I didn't get involved anymore.
It's only a matter of time before Sienna gives three AI teams the task of organizing their work.
I took up the caramel sugar nitrate in my hand, ready for the siege.
It's just that I've never had a cup of coffee.In less than eight minutes, the three young AI writers came out with freshly-boiled questions and credentials, continuously reaching back from all sides to cross over to the main writer.
More interestingly, Sienna is still a bit of an editor.
After receiving the selection, it does not directly take a complete picture, but checks the quality step by step.
If you hit a blurred source, you can't find a link to report it, and it'll list a complete list of receipts directly, and I'll alert the corresponding agent as soon as I approve.Watch the standards, find them again!。
So, as soon as I was busy, I started the task and looked at it.
Oh, my God!
A summary of the qubit topics containing six of the latest AI dynamics has been put together in a neat way:
Candidate headlines, news dates, core facts, reasons for recommendations, links to credentials are all missing.
Even three AI writers were attached to their respective assignments, reminding me of the details that still need to be retested.
When the briefing was delivered with tears in my eyes, I suddenly raised a new suspicion and interviewed the agent who was in charge of the staff of the big factory.
Where the hell is Tibo going?
As a result, it didn't ask me who Tibo was, but continued to look directly at previous human clues, and soon found that Tibo's timing and public moves toward shaving his head were straightforward.
I've even handled the idea of writing a script that doubles the effect of a show.
If you put Tibo and Cat Wu, who recently jumped into Tibo's hand, together, the effect would be even better!
That's good, that's good, that's good.
Not only can I send, receive, review, and temporarily activate it, but I can even combine it with my small role as a writer to initiate planning.
These AI editors are a bit off the hook.
Test Two: Just after the briefing, the next great web page is already connected, and it can even attract people on the spot.
After completing the exam, I temporarily sent a new one to Sienna.
This time, I want it to be able to turn the short summary of the new question into a resource that can be filtered, searched, and jump-started.AI news site。
And it is not allowed to reconnect to find new ones, which must follow the results of the previous round of elections.
I thought Sienna would bring back the three original AI writers to work together.
As a result, it is the first to grind itself.
The expertise of human resources writers can't play around with web development, they just don't call it that anymore.
While there is so much relevance to finding large models and open source projects, there is also some professional disagreement about how much is used.
What about those who are temporarily missing?
That's a real trick!
(I was so surprised that I didn't even find out until after the mission, when I opened the side mission box...)
After roughly 10 minutes of grinding, Sienna directly built two new Workers on the site, one of which was called the "Workers".Quotation summary of web development of quotationIt is responsible for the creation of web pages.
The other is calledQQQQQQQQQQQQQSpecializing in quality testing and acceptance.
And also,The new Worker connects directly to the previous round's six topics.The department began to complete content preparation, web development and QA acceptance.
I don't need to re-write the material at all.
It's going to happen soon.The structured documentation of six pieces of information was delivered first, and four inspections were all approved.
Afterwards, the web development worker immediately knocked on the code, and QA simultaneously raised two questions.
The development of Worker received feedback and was decisively modified.
In the end, an AI news hotspot was successfully boiled and took about a year to complete.Thirty minutes.
Believe me, I really didn't think...
Sienna has already begun to deploy himself according to the needs of the mission.
What's more, it also knows when to change people and how to get new colleagues to pick up the pieces.
And from beginning to end, I haven't heard a new command except to read it.
The computer is off, why are the Agents still working?
Seeing here, maybe some of your friends would like to ask.
Why is Sienna a newly hired Worker who can take the next round of work directly?
The key is:LobeHub puts work goals, task progress, delivery files, and acceptance records into a shared workspace.
The managing agent can arrange follow-up tasks around the results, and the worker who takes over knows what has already been done and what needs to be done next.
So I don't have to repeat the six selections again, and a new round of work has already begun.
So how is it possible for these agents to be able to deploy themselves, temporarily recruit, or even inter-project?
The answer lies in the LobeHub operating system.
First of all,The agent's work is no longer confined to the chat room.
LobeHub de-manages the Agent Loop, which is responsible for continuous execution of tasks, the Session, which saves the status of tasks, and the operating environment for calling tools.
In this way, the user withdraws from the conversation and the agent does not have to follow the interruption.
As long as the mission operates in a sustainable environment, it can continue to advance in the background and maintain previous work progress.
The LobeHub team also used a client data migration task to test water:
This group of Agents is running fast.2 days 3 hours 47 minutesIt was demolished.60 of them.Mission accomplishedFifty-six timesIndependence and migration.
Two days of sleeplessness... Six, the AI bull horses are really doing their job this time...
△Long-term Goal execution recordsThere's not enough light to keep working.It's just a matter of connecting the different agents.
For this purpose, LobeHub organizes work goals, mission progress and delivery results so that the receiving agent knows what the previous colleague has done and what to do next.
And then there was the scene before it:
I've just finished six subjects, and the second round will take the completed summary to create an AI hotspot page.
Of course, it can be delivered live, and the quality can't keep up with the spread.
LobeHub is also equipped.Acceptance Multimodal Acceptance。
In addition to the final results, the agents will be assigned to work with evidence such as test results, screenshots and recording screens.
So we can look directly at the actual effect, find the problem, go straight back and modify it, and be satisfied and confirm the approval.
As I saw in my second round of testing, the web code, although submitted, was not fully validated and could not be directly calculated as having completed the acceptance.
This means:In the workflow of the LobeHub, human beings have always been firmly in control of the final acceptance, despite the fact that the Agent is a fast-acting individual.
△Acceptance Acceptance MechanismSo a set that could both enable the Agency team to work and allow humans to control the workflow of the acceptance authority, is running.
From LobeChat to LobeHub, they've built AI into the company step by step.
In fact, LobeHub didn't learn how to make an AI team work overnight either.
As early as 2023, they were open source AI chat tools.LobeChat。
LobeHub's AI has steadily moved out of the chat room with the addition of capabilities such as Agent, Tool Calling, and Knowledge Base.
In January 2026, LobeHub 2.0 further introduced features such as Agent Group, Memory and more to allow multiple Agents to start team collaboration.
In May, they launched a new one.CAO (Chief Operating Officer of Intelligences)The ability to organize, dismantle, activate, and coordinate tasks with other Agents began to be handed over to AI.
It's a coincidence.Heroes see little in common。
This idea coincides with the recent heated debate.DeepSeek Super-Agent architecture is exploredIt's not worth it.
Half a century ago, LobeHub turned collaboration into a product, and now it's open source so that other developers can deploy the experience themselves.
What's even more interesting is thatIt's the company's first knife.
Self-made Agents, who in turn are involved in the development of their own products, from writing code, auditing PR to cross-project connectivity, are truly integrated into everyday R&D.
LobeHub has also put forward a more audacious goal as AI colleagues look forward to doing more and more.Self-Driving Organizations (self-driving organizations)。
According to their proposed L1 and L5 ratings, the team is currently operating in phase L3, with some scenes being directed to L4s that are able to autonomously send, receive, and accept L4s.Autonomous business unitLet's move.
△Self-Driving OrganizationsAnd in retrospect, my observations are also interesting.
LobeHub developed LobeHub with LobeHub;
I actually used it to develop a quantum AI editing department.
Two completely different jobs, ultimately running on the same set of collaborative logic.
This may be the next stage of AI Native organization.
AI is responsible for making things work, and people are responsible for deciding where to go.
The result was a 1+1>2 effect.
Anyway，LobeHub has put the source code for this AI collaboration platform on GitHub so that interested partners can deploy the experience themselves.
We also welcome you to try it yourself, come to the comment area and share with us what your AI colleagues have created!
(P.S. Mommy honestly said that the information I needed to write this article was basically the stuff LobeHub helped me sort out.
The experience addresses are as follows:
GitHub directions:
