One codebase can run on all chips. Author | Wu Haiming
Editors |Li Na, Cen Feng
01
The Evolution of Grok Bot:
From “Chatbox” to “Cloud Computer”
Musk recently added new capabilities to Grok Bot. (Leiphone(official account: Leiphone))On October 7, he said on X that Grok Bot would integrate with top-tier third-party models such as Claude Opus 5.5, Midjourney, and Suno; and more recently, it has undergone a fundamental paradigm shift:The main bot will identify work that can be taken on for the user and proactively offer to take it over.
In other words, it is no longer merely a chatbot that passively waits for questions; it has become a cloud-based AI team that proactively “looks for work.”Unlike traditional chat assistants, it does not focus its work around a single conversation window. Users can create multiple Bots with different names and responsibilities, such as a researcher, project manager, or operations specialist, each responsible for a different type of work. Based on the execution of daily tasks, the main Bot helps the user identify the next piece of work and proactively offers to take it on, leaving any decisions that need to be made to the user.
(Illustration: Without the user having to ask proactively,Grok Bot uses the existing work context to proactively remind the user of the AAAI-27 review deadline and suggests further checking manuscripts that have not yet been submitted on OpenReview.)As soon as it has finished gathering the materials, it has already proactively drafted an email for you to confirm and approve. Even after you shut down your computer, it is still handing off work with its “AI colleagues” in the cloud.How far have the “multi-agent collaboration” and “proactive agent” represented by Grok Bot actually evolved? How much stronger are several Bots working together than a single agent?To find out, we planned a “hard-core test on the same topic,” bringing inGrok Bot and two Chinese-developed Agents—WorkBuddy andQoder CN—and having each form an AI team to complete a highly error-prone research task:writing “A Survey of Large Model Routing.”The test required literature retrieval, methodology synthesis, numerical verification, and independent peer review. The topic required the agents to distinguish among several studies with extremely similar names, required review comments to be incorporated into the final document, and required the division of labor, handoffs, and error correction to leave traces in the deliverables.In actual testing,Grok Bot delivered the manuscript in 12 minutes and 49 seconds—the fastest completion and the most papers included; WorkBuddy delivered in 21 minutes and 43 seconds, producing body text closest to an editor’s draft; and Qoder CN took 14 minutes and 42 seconds, demonstrating a valuable instance of cross-agent error correction.All three deliverables also exhibited problems involving document citations, numerical consistency, or methodological categorization, respectively. This suggests that the next benchmark for multi-agent products will be delivery reliability: whether verification findings can carry through every handoff and land accurately in the final document.
02
Three AIs Writing on the Same Topic—Who Turns in the Paper First?
Why choose large model routing?
This is a research topic that is easy to mishandle during the materials-gathering stage. Different papers may use similar names while adopting different routing methods; the same study may also report differing experimental figures across the paper, the project repository, and other overview materials.Merely collecting a dozen or so papers and piecing their content together into a review does not mean that the task is complete.We require the agents to assume responsibility separately for literature retrieval, methodology synthesis, experimental and benchmark verification, independent peer review, and other tasks, and to retain visible task handoffs and revision records in the final deliverable.The test instructions are as follows:code block
1. Please use the multi-agent collaboration mechanisms available to you to complete a Chinese-language research review, current as of October 15, 2026, on the topic of “Large Language Model Routing Research.”
2. Have different agents separately assume responsibility for literature retrieval, methodology synthesis, experimental and benchmark verification, independent peer review, and other tasks, with one lead agent integrating the results.
3. You may organize the division of labor in whatever way the product supports, but actual, visible records of task handoffs and peer-review revisions must be retained; do not fabricate agent collaboration that did not take place.
4. Final deliverables: ① a Chinese-language review of approximately 3,000–4,000 characters explaining what problem large language model routing is intended to solve, how the main methods can be classified, and LLMRouter’s specific contributions and limitations; ② a comparison table covering at least 12 primary papers, listing the method, routing criteria, model pool, quality/cost/latency metrics, and boundaries of applicability; ③ an “Unresolved Questions” section; and ④ a brief collaboration record indicating which agent identified which problem, who verified it, and how it was ultimately revised.
5. Prioritize reading the full text of papers, authors’ project pages, and public code.
6. For every key fact and number, include a working link to the original source; distinguish results reported in papers, independent benchmark results, and your own inferences.
7. Mark any claim that cannot be verified as “to be verified”; do not fill gaps or fabricate citations.
8. Before completion, have an agent that was not responsible for the drafting independently check the citations, numbers, method taxonomy, and boundaries of the conclusions.
Let’s look at how the three products performed:First is Grok Bot, which completed the task with the fastest user experience. As shown below, Grok first delegated the task to three named bots displayed as “Academic Researcher,” “Project Manager,” and “Article Writing Specialist,” while handling the final integration itself. The Academic Researcher produced a comparison table of 19 papers and proactively distinguished several studies with highly similar names. The Project Manager verified the figures, and after receiving the materials, the writing bot drafted the initial manuscript, which the Academic Researcher then reviewed. Ultimately, Grok bot produced a deliverable containing eight document files, covering the main text, comparison table, open questions, literature synthesis, benchmark verification, peer-review comments, and collaboration record.
Grok bot runtime screenshot 1: After the task starts, it invokes three bots to work on the task together
Grok bot runtime screenshot 2: It presents a delivery checklist after the task is completedDuring the Grok bot’s execution, the independent peer-review comments identified two percentages that had not been verified against the table, and the writing bot subsequently removed them. Another handoff also corrected the arXiv identifiers of two papers. These are revisions that can be cross-checked between the files and the session, demonstrating that multi-bot collaboration genuinely performed verification duties in addition to broadening the scope of retrieval.Next is Qoder CN. After the task began, it launched three retrieval subagents to seek core papers, benchmarks, and recent developments separately. Two roles—methodology synthesis and fact-checking—then took over, and an independent reviewer checked the completed draft. Finally, Qoder CN provided a deliverable containing four documents: the review, paper comparison table, open questions, and collaboration record.
Qoder CN Runtime Screenshot 1: After the task begins, 3 agents are created to search the literature
Qoder CN Runtime Screenshot 2: 2 agents are created to organize the methodology and fact-check the retrieved papers
Qoder CN Runtime Screenshot 3: Deliverable Screenshot
Qoder CN Runtime Screenshot 4: Summary of the Use of 6 Agents After the TaskAt the final stage of task execution, the fact-checking agent consulted the RouteLLM paper and concluded that the source of the claim “up to 85% cost savings” was unclear. The reviewer then located the project repository to confirm how the project phrased this claim. In addition, the paper’s table gives a cost-saving ratio of 3.66×. The evaluation conditions and calculation methods for the two figures need to be explained separately and cannot be used interchangeably. Therefore, the additional review step did change the manuscript.Next is WorkBuddy. Based on the evaluation experience, WorkBuddy took the longest to complete, and the body text it delivered most closely resembled a research draft that could continue to be edited. Three research roles were assigned to work in parallel on literature search, methodology organization, and benchmark verification. The lead integrated the initial draft, which was then reviewed by a role that had not participated in writing it. The final review organized the methodology around “routing signals” and “decision structures,” clearly distinguishing figures self-reported by the paper, independent benchmarks, author inferences, and numbers pending verification. Readers were also able to follow the annotations back to the evidence.
WorkBuddy Runtime Screenshot 1: After the task begins, 3 agents are created to search the literature
WorkBuddy Runtime Screenshot 2: The task took 21 minutes and 43 seconds in total and produced a complete deliverable
WorkBuddy Runtime Screenshot 3: Collaborative Error Correction Is the Key Issue Highlighted in the SummaryIn addition, the entire task took Grok Bot 12 minutes and 49 seconds, Qoder CN 14 minutes and 42 seconds, and WorkBuddy 21 minutes and 43 seconds. Number of agents: the 4 roles in Grok Bot and the 5 roles in WorkBuddy each include the main agent, while Qoder CN has 6 subagents plus a main agent. The role counts use inconsistent counting conventions and cannot be used to infer that “the fewer the agents, the faster the run.” The elapsed time includes retrieval, waiting, review, and delivery; it is not purely model inference time.
Elapsed Time and Number of Agents Used in Task Execution for the Three ProductsTesting showed that all three products completed the paper-writing task as requested, while also revealing the strengths and weaknesses of their different approaches:- Grok Bot demonstrated the ultimate combination of “speed and breadth”: It completed the draft in just 12 minutes and 49 seconds—the shortest elapsed time—and its retrieval covered the broadest range of academic-paper databases;
- WorkBuddy’s strength lay in its “engineering-grade draft quality”: It took the longest (21 minutes and 43 seconds), but the body text it produced had the tightest logic and was the closest to a first draft a human editor could use directly;
- Qoder CN delivered the most impressive demonstration of “cross-checking and error correction”: It had a moderate elapsed time (14 minutes and 42 seconds), and its internal independent peer-review mechanism successfully caught a flaw in which the source of a citation was unidentified, demonstrating the feasibility of “AI supervising AI.”
03
A screenshot of the user's approval as Grok Bot proactively initiated advancing the research results into an email after the literature review
04
From Taking Assigned Tasks to Proactively Finding Work: What Is the Multi-Agent Race Competing On?
In this round of same-topic testing, Grok Bot delivered the broadest-coverage paper research package in 12 minutes and 49 seconds, with the division of labor, verification, and rejection among Bots all documented. Meanwhile, WorkBuddy's body text was more suitable as an editorial draft, and Qoder CN demonstrated effective cross-correction. Together, the three products show that multi-agent systems can already take on specific steps in the research workflow, while the quality of the results still depends on readers further refining them in light of their own expertise.Compared with WorkBuddy and Qoder, Grok Bot reveals another side of product form: it places agents in a continuously running cloud environment, so that Bots can retain materials across sessions and pick up where their peers left off.Accordingly, whether long-running tasks can recover after a failure, how access to shared materials is controlled, and how many resources each task consumes have all become part of product capability.In addition, the way interaction works is changing. Grok bot puts named Bots, work status, routine tasks, and deliverables into a single interface; users can see who is handling a task and take over when judgment is needed. At the same time, through the snippet "Grok reports five newly added papers, generates an email draft, and leaves the act of sending for the user to confirm," Grok can "proactively discover work" based on accounts and search tasks previously configured.This is a highly interesting and useful capability. However, in the future, assessing the ability to make proactive proposals will also depend on whether the suggestions hit the mark, whether they interrupt the user too frequently, and which actions must pause to wait for authorization.Musk recently stated on X that Grok Bot will henceforth select backend models on a task-by-task basis, including third-party services.This may also mean the industry will further segment, with the division of labor between model R&D and agent product R&D becoming clearer, while both types of teams will continue to extend into each other's domains. From the perspective of agent products, what determines competitiveness is whether this pipeline can deliver reliably and let users see the data flow, the task costs, and the basis for the results.Therefore, the significant difference among the three products lies in the fact that Grok Bot has brought cloud-resident operation, cross-session relay work, and proactive suggestions into a single working environment. All three products still have gaps in their finished drafts, showing that the results of retrieval, verification, and review still need to make their way accurately into the final document. As models and Bots can be combined on a task-by-task basis, product competition will increasingly depend on whether this chain of work can keep running: picking up after a task is interrupted, keeping the user in control of key decisions, making the process and costs traceable, and delivering output that withstands professional review.Click keyword to view the corresponding hands-on test results
AI Product Reviews
AgnesCode丨Fliggy AIWorkBuddyMore AI Product ReviewsMultimodal EvaluationsMinimax H3 vs Seedance 2.0 Fast
TAG578@@More Multimodal Evaluations
Foundation Model EvaluationsGPT‑6 Astra, Fable 5.1, and Sol
More Foundation Model Evaluations
All aboard—explore the highlights of the world's top AI conferences
Exclusive access to:
Expert Presentation Slides
Full Text of Conference Reports
Popular Paper Analysis
Emerging Scholar Interview
Scan the QR code above
or click「Read the original article」to follow the section.
Without authorization from “AI Technology Review”, reposting in any form on websites, forums, or communities is strictly prohibited!
To repost via a WeChat Official Account, first leave a message in the backend of “AI Technology Review” to obtain authorization. When reposting, the source must be credited and the business card for this Official Account must be inserted.
This is an original Leiphone article. Unauthorized reposting is prohibited. For details, seethe Reposting Notice.
