DeepSeek: Here Come the Super-Agents

DeepSeek Publishes “Super-Agent” Concept on GitHub

DeepSeek has introduced the concept of a “super-agent” designed to coordinate multiple AI agents autonomously and largely remove humans from supervision. The initiative is based on a technical paper released over the weekend by the internal DeepSeek-Harness team titled “Exploring Super-agent Design: Towards a More Autonomous Agent System.” The company argues that improved models have made agents viable for an increasing range of tasks, from calendar management and email drafts to work reports and coding. The bottleneck, it claims, now lies with humans, who must distribute tasks, monitor progress, and resolve conflicting outputs.

According to the team’s assessment, an individual can typically supervise only three to five active agent sessions simultaneously. This limits how much work organizations can delegate, even when individual agents handle their tasks independently. The super-agent is designed to push past this limit via a centralized coordination layer. It breaks down larger jobs, assigns the subtasks to parallel agents, and tracks their progress without requiring a human to manage each session individually.

The system is also designed to react to events rather than waiting for explicit prompts. As an example, DeepSeek cites a bug report posted in a group chat: the agent attempts to reproduce the bug, develops a fix, and, if successful, submits it for human review. If the attempt fails, the agent remains silent, avoiding unnecessary notifications.

Alongside the paper, DeepSeek released an MIT-licensed prototype named dsh-superagent on GitHub. It is a proactive coding agent designed for a single GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and optionally the messaging platform Feishu. When an authorized user requests a task, the agent initializes a workflow, executes it in dedicated sessions and directories, and replies within the original thread, for instance by opening a Pull Request. A web dashboard indicates where human intervention is required. The team formulated a core guiding principle of demanding as little human attention as possible: the agent acts only on explicit request or according to predefined team rules.

DeepSeek explicitly labels the code as a research prototype rather than an official product. It differs from both the internal version and the implementation outlined in the paper. Several features described there represent forward-looking research directions not yet included in the public repository. According to the team, the prototype emerged directly from day-to-day engineering workflows, with a substantial portion authored by coding agents themselves. Even the installation is handled by an agent that walks users through the repository’s setup instructions step by step.

With this initiative, DeepSeek—built by Liang Wenfeng out of the quantitative hedge fund High-Flyer—is expanding from open models into the agent execution layer. There, the company directly challenges Anthropic’s Claude Code and OpenAI’s Codex, both of which already operate across full codebases to execute tasks independently. Anthropic has reported an annualized run rate exceeding $2.5 billion for Claude Code. → South China Morning Post, RuntimeWire, DeepSeek on GitHub, TechOrange 科技報橘

Synthszr Take: DeepSeek is giving away the blueprint before selling a product. The prototype is MIT-licensed and explicitly marked as research, meaning it promises nothing and binds no one. Yet it runs exclusively as a bundle for DeepSeek Harness. Every developer experimenting with it absorbs DeepSeek’s orchestration logic. Claude Code pulls in more than $2.5 billion annualized. Against that massive lead, DeepSeek is using a weekend paper to buy a seat at the table defining the agent control plane.

OpenAI Seeks to Rebrand Data Centers as “AI Compute Campuses” After Midterms

OpenAI plans to launch a political and PR campaign immediately following Tuesday’s US midterms to adapt to a shifting Washington landscape, report Jim VandeHei and Mike Allen at Axios. The strategy stems from expectations that Democrats will win the House of Representatives, subjecting the AI industry to increased scrutiny after two years of favorable treatment from the White House. According to Axios, OpenAI’s Chief Global Affairs Officer Chris Lehane and his team have been preparing new and repackaged policy proposals since August, slated for introduction when Congress convenes its Lame-Duck-Session on November 9. A cornerstone is mandatory safeguards for premier models, which Lehane calls “speed limits,” insisting that humans must retain ultimate control over the technology. To counter bipartisan local backlash against data centers, OpenAI intends to offer community benefit agreements and advocate for state-level legislation granting residents a voice in project construction and dividends. → Axios

Synthszr Take: OpenAI is writing the playbook before the incoming Congress gets a chance. The “speed limits” will land during the lame-duck session starting November 9—pitched directly to lawmakers from the outgoing majority. Any Democratic majority seeking tighter regulation come January will find itself negotiating against an industry standard already on the record. Imposing stricter mandates will demand real political capital. For OpenAI, this entire maneuver merely required funding a policy team since August.

Mistral Raises €3 Billion and Makes Its Money from Projects

Parisian AI startup Mistral is quietly pivoting its business model, according to an investigation by manager magazin. The firm is increasingly deriving revenue from infrastructure and enterprise project services rather than selling proprietary foundation models. Mistral recently secured €3 billion in fresh capital at a valuation of €21 billion, setting a record for a private European tech enterprise. Publicly, founder and CEO Arthur Mensch (34) maintains the ambition to lead the technological frontier. Manager magazin identifies the compounding technological lead of US labs and brutal price competition triggered by Chinese open-weight models from providers like DeepSeek as key drivers behind the pivot. → Tech Update – manager magazin

Synthszr Take: Two trillion dollars is, above all, the cost of scarcity. Much like Arm in years past, public markets lack direct avenues to bet on a frontier AI lab. Consequently, fund managers pay exorbitant premiums just to gain entry. They receive no governance rights in return, as class-F shares keep 50.1 percent of voting control firmly in the founders' hands. Mistral reveals how razor-thin model margins truly are: Europe’s most valuable tech startup earns its keep configuring data centers for Microsoft and embedding consulting staff on site.

Claude Code Refuses Recreations of Popular Games, Developers Report Blocked Projects

Anthropic is reportedly clamping down on its coding agent Claude Code to prevent it from rebuilding existing video games. According to VideoCardz, developers have reported stalled projects after the tool rejected requests visibly intended to replicate existing titles. This constraint disrupts ongoing projects that can no longer proceed using Claude Code. The exact wording and scope of the restriction have not been publicly disclosed. → VideoCardz.com

Synthszr Take: Guardrails in one tool cannot protect a game when the next tool happily accepts the prompt. Cursor allows engineers to switch between multiple models, rerouting a rejected prompt in a single click. The true verdict will emerge in competitors' adoption metrics over coming weeks. If indie developers leave with their half-built clones, Anthropic will have succeeded primarily in curbing its own liability while gifting paying customers to rivals. If they stay, Anthropic has drawn a boundary the market is willing to accept.

Prompted by AI Detectors Like Pangram, Writers Ditch Em Dashes and Insert Typos

Writers are actively altering their prose style to prevent AI classifiers such as Pangram from flagging their work as machine-generated, according to a recent report. Pangram is widely used to evaluate whether text originated from a large language model. The report highlights that authors increasingly shun the em dash, now commonly treated as a hallmark of synthetic output. Stylistic structures following the “not X, but Y” formulation are similarly being discarded. → Wired

Synthszr Take: Pangram shifts the burden of proof onto the author. Detection models evaluate probabilities, where em dashes combined with words like “delve” inflate synthetic probability scores. The victims are precisely those writers with a distinctive voice, as their idiosyncratic quirks mathematically resemble the statistical patterns of frontier models. Publishers, agencies, and universities relying on automated filters inevitably receive sanitized, homogeneous prose. The cost is unique personal style, leaving the deliberate typo as the final badge of human authenticity.

Apple Brings Apple Intelligence to China, Powered by Alibaba’s Qwen

Apple is rolling out its Apple Intelligence suite in China by utilizing Alibaba’s Qwen models rather than its in-house systems, reports TechOrange. Following the October 31 announcement, a developer beta is scheduled to deploy on November 4, ahead of a public rollout planned for December with iOS 27.3. The report notes that Qwen models will be integrated into Siri and Writing Tools. The partnership addresses Chinese regulatory requirements mandating that all commercial models pass local compliance reviews, an obstacle that had kept Apple Intelligence absent from the mainland market. → TechOrange 科技報橘

Synthszr Take: In China, Apple is compromising on the foundational hook it uses to market Apple Intelligence across Western markets: end-to-end proprietary privacy. The moment user queries route through Alibaba’s infrastructure, Private Cloud Compute turns into a promise with local asterisks. The compromise remains economically imperative: in Apple’s second-largest regional market, domestic manufacturers are winning share with aggressive on-device AI features, and an iPhone without an upgraded Siri risks lost sales during the critical holiday quarter. In exchange, Alibaba secures what Google pays billions to maintain in Cupertino: permanent real estate inside the operating system.

Anthropic Uses More AI Internally at List Price Than It Pays in Staff Compensation

According to Business Insider, the value of internal AI usage at Anthropic now surpasses the total compensation package of its entire workforce. The metric tracks Claude tokens consumed by employees in their day-to-day operations. The calculation assesses this internal usage at commercial retail rates—the exact fee an external customer would pay for equivalent volume. This benchmark reflects the open-market value of the compute, while the company’s actual overhead consists solely of the raw cost of Inference. → TechOrange 科技橘

Synthszr Take: Anthropic pays only inference compute costs for these tokens, whereas clients pay retail list price. That spread creates an insurmountable margin advantage that no outside software vendor can replicate. A startup using Claude with matching intensity spends more on tooling than on payroll. Anthropic acquires equivalent labor output at pure wholesale cost. As software development transitions into fully agentic workflows, this pricing delta will dictate competitive margins and execution speed, particularly against enterprises attempting to build on Claude’s stack.

New York Times: Suspected Chinese Agent Was Close to Nvidia CEO Jensen Huang

US intelligence agencies reportedly concluded that Alice Huang, an executive at Singapore-based data center operator Megaspeed, passed intelligence to Beijing while maintaining close personal ties to Nvidia CEO Jensen Huang, according to The New York Times. President Trump was reportedly briefed on the matter approximately one year ago without the White House taking administrative action. The precise nature and intelligence value of the material transmitted remains unknown. Officials were concerned she could provide visibility into Trump’s policy positions on semiconductor exports to China as well as proprietary insights into Nvidia. Intelligence officials assessed that Jensen Huang was entirely unaware of her alleged role, though he was never administered a Defensive Briefing. Megaspeed is already subject to US federal scrutiny over allegations of rerouting Nvidia chips to China. Via an affiliated entity, it procured at least $4.8 billion in Nvidia hardware and maintains an active cloud partnership with the chipmaker. Nvidia dismissed the engagements as routine, stating that enterprise customers receive “zero information relevant to national security.” → The Rundown AI

Synthszr Take: The White House sat on the file for an entire year, yet Jensen Huang was never granted the standard defensive briefing routinely provided to US executives under surveillance. This clearly illustrates Washington’s geopolitical priorities: Nvidia’s silicon serves as diplomatic leverage against Beijing, with counterintelligence taking a back seat. Megaspeed remained a recognized cloud partner despite ongoing probes, backed by more than $4.8 billion in capital expenditure. For enterprises sourcing compute through Southeast Asian intermediaries, the route remains hostage to geopolitical calculations—and a single press report can sever it overnight.

State-Funded Study Warns Against AI in China’s Social Credit System

A study sponsored by the Chinese government warns against integrating AI into the nation’s social credit framework, cautioning that it risks turning into a “large-scale instrument of surveillance.” Authored by researcher Lin Jianxing, the paper was published in the Journal of Sichuan Normal University and funded via a state-backed research initiative on social governance and institutional innovation, Interesting Engineering reports via the South China Morning Post. The existing system does not function as a single national score, but rather as an aggregate of commercial credit metrics, bureaucratic blacklists, local pilot programs, and disciplinary tracking systems for civil servants. Lin cautions that AI could evaluate behavioral trends, financial histories, and interpersonal networks at scale. Citizens could subsequently be classified based on statistical similarity to existing “problem individuals,” absent any statutory infraction. A 2025 Stanford examination of a local system where adults start at 1,000 points subject to 389 behavioral rules documented automatic deductions for filing administrative petitions, as well as structural disadvantages for rural citizens.

Synthszr Take: The fragmentation of the system has historically been its sole protective safeguard. Banking data, administrative blacklists, and municipal scoring systems rarely interoperate, and every institutional gap leaves room for appeal. AI bridges those siloes. A score derived from algorithmic similarity identifies no discrete legal violation that could be challenged in court. This directly penalizes those who are already marginalized. In a localized framework containing 389 strict rules, merely lodging a petition with local officials costs points; analyzed through predictive clustering, a legitimate grievance becomes a registered risk profile.

Meta Builds Household Robot Max in Three Months with Chinese Partner Fourier

Meta has developed a humanoid robotics prototype engineered for domestic household tasks, according to a report from The Information. Internally designated Max, the machine was constructed in approximately three months alongside Chinese robotics manufacturer Fourier. According to sources, Max weighs roughly 48 kilograms and stands 1.5 meters tall. The prototype reportedly features an integrated display for facial expressions, internal speakers, and dexterous hands capable of manipulating small items. For onboard processing, Meta utilizes Nvidia’s Jetson Thor computing platform. → TechOrange 科技報橘

Synthszr Take: Meta absorbed an expensive lesson during the smartphone era. Apple compromised its core advertising engine with a single iOS privacy update because Meta was merely an application running on someone else’s hardware. Zuckerberg has no intention of repeating that vulnerability in physical robotics, leading him to engineer the hardware directly. Yet the dependency has merely shifted: Max was built in ninety days with Fourier and computes on Nvidia’s Jetson Thor. For now, in-house hardware simply amounts to a proprietary casing built around an external supply chain.