LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-03100 Beiträge

Aktualisieren

CoreWeave and Convergent Science Complete 536 Million Cell Direct Numerical Simulation on NVIDIA HGX B200

Kurzfassung
2026-11-03 00:34 GMT+8

Direct Numerical Simulation (DNS) is the most accurate method for simulating turbulence in computational fluid dynamics (CFD), solving all scales of fluid motion without relying on turbulence models. However, its extreme demand for memory and compute resources often limits its application to smaller cases or requires specialized high-performance computing infrastructure.

CoreWeave, an AI infrastructure provider, partnered with Convergent Science, a CFD software developer, to benchmark CONVERGE CFD software on NVIDIA HGX B200 GPU clusters. The test successfully completed a DNS case containing 536 million grid cells in 3.62 hours using two NVIDIA HGX B200 instances, comprising 16 NVIDIA Blackwell GPUs.

The study also compared two distinct hardware classes: NVIDIA RTX PRO 6000 Blackwell Server Edition and NVIDIA HGX B200. The RTX PRO server is suited for everyday industrial simulations that fit within a single node, utilizing PCIe for intra-node communication. In contrast, the HGX B200, with significantly higher memory bandwidth (~7.7 TB/s vs 1.6 TB/s) and NVLink interconnects, is designed for ultra-large cases spanning multiple nodes.

It is important to note that these results come from a vendor-partnered benchmark designed to demonstrate theoretical upper limits and scaling characteristics under specific conditions. Real-world production performance may vary depending on physical models, boundary conditions, and software optimizations.

Quellen:wf.coreweave.com

Forschung

South Korean Fintech Unicorn Toss Plans Acquisition of Japanese Brokerage Cher Securities

Wesentlich
2026-11-03 06:58 GMT+8

According to a report by South Korean media outlet Byline Network, Viva Republica, the operating entity behind the fintech giant Toss, is planning to enter the Japanese securities market. The company intends to acquire 100% of the shares of Cher Securities, a mobile-specialized brokerage under the Tokai Tokyo Financial Group, through its overseas business arm, Toss Global. The transaction value is estimated at approximately 4 billion yen (around 34 billion won), with a target completion date by March 2027.

Cher Securities, established in 2019, is an online brokerage focused on small-scale stock trading and investment trust services via smartphones. Toss stated that the decision to acquire Cher Securities stems from the alignment of their service philosophies, particularly the focus on accessibility for beginner investors. The strategy involves leveraging Cher’s existing customer base and operational experience while integrating Toss’s mature mobile trading system (MTS/WTS) technologies to gradually enhance services without aggressive restructuring.

This strategic move reflects Toss’s search for new growth engines. While Toss boasts a massive user base in South Korea, its profitability relies heavily on its domestic securities business, where growth margins are tightening. Compared to the fierce competition and strict regulatory barriers faced by foreign entrants in Japan’s banking sector, the securities industry offers a more realistic entry point. Japan’s relatively stable political environment and well-established financial regulatory framework are also viewed as favorable conditions for testing digital financial platform models.

For investors, this acquisition represents a critical step in Toss’s globalization strategy. Previously, Toss launched its eponymous app in Australia via Toss Global. If the integration in Japan proceeds smoothly, it could contribute direct overseas revenue and add a significant “cross-border platform capability” narrative to the company’s potential IPO valuation. However, the deal is not yet finalized and remains subject to approval by the Japanese Financial Services Agency (FSA) and subsequent integration risks.

Quellen:바이라인네트워크

Forschung

Anthropic Updates S-1: Q3 Revenue Hits $17 Billion as Operating Losses Widen Ahead of IPO

Thema · Anthropic上市Strukturell
2026-11-03 07:19 GMT+8

U.S.-based AI company Anthropic published an updated version of its S-1 registration statement for its upcoming initial public offering on November 2, 2026. The document reveals that the company generated $17 billion in revenue during the three months ended September 30. This figure represents 11.27 times the revenue from the same period a year earlier, but indicates a significant deceleration in growth: the quarter-over-quarter increase was 47%, down from 144% in the previous quarter.

Despite the substantial revenue, Anthropic reported a $536 million operating loss in the third quarter, compared to $974 million in operating income three months earlier. The primary driver of this swing was a $5 billion charge incurred because demand for computing hardware exceeded expectations, forcing the company to purchase equipment at a premium. Additionally, the company took a $2 billion charge related to a non-cash employee donation matching program. Excluding stock-based compensation and charitable contributions, adjusted operating income rose by $300 million quarter-over-quarter to $2.5 billion.

The S-1 also details Anthropic's massive infrastructure commitments. As of September 30, the company had $895.4 billion in outstanding infrastructure commitments. Leases signed with Google LLC and Broadcom Inc. accounted for nearly a third of this sum. Anthropic plans to not only rent more computing capacity but also build its own data centers, with purchase commitments jumping from $55 billion at the end of 2025 to approximately $314 billion in the third quarter.

Anthropic has not yet disclosed how much it plans to raise through its public offering. According to Bloomberg, the company hopes to at least match the $86.2 billion that SpaceX raised when it listed its shares earlier this year. Reuters previously reported that Anthropic is seeking a valuation of $1.5 trillion to $2 trillion.

Quellen:SiliconANGLE

Forschung

McLaren Deploys Physics AI and Agents to Cut Vehicle Design Evaluation from Days to Hours

Thema · 迈凯伦AI工程平台Wesentlich
2026-11-03 06:25 GMT+8

British high-performance sports car manufacturer McLaren Automotive has announced a partnership with engineering computing platform Rescale and chipmaker NVIDIA to integrate Agentic AI and Physics AI into its vehicle engineering development process. This initiative aims to address the core bottleneck of traditional Computer-Aided Engineering (CAE): high computational costs that limit the scale of design exploration.

In conventional automotive engineering, evaluating body structure stress, dynamic performance, and material durability requires running complex numerical simulations. While these high-fidelity simulations provide reliable physical evidence, each run consumes significant time. When engineers need to repeatedly adjust parameters or compare numerous design options, computational resources become a constraint, limiting the number of designs assessable within the development cycle. For high-performance vehicles requiring precise tuning, accelerating iteration while maintaining analytical reliability is a critical challenge.

McLaren's solution combines "Surrogate Models," a type of AI that learns physical system behavior from existing simulation or experimental data to map relationships between design parameters and physical outcomes. Instead of re-running high-precision numerical simulations every time, surrogate models can predict results under different design conditions at millisecond speeds within their applicable range. According to data provided by Rescale, McLaren engineers can now evaluate thousands of design variations in hours, compressing complex simulation tasks that previously took days into minutes. This mechanism allows teams to first screen out designs worth deeper analysis using low-cost predictions, then perform high-precision validation on key candidates.

Beyond Physics AI, the platform introduces AI Agents to handle repetitive engineering tasks. Rescale's platform integrates agentic engineering technologies into McLaren's R&D environment, automating the chaining of simulation and analysis workflows. This frees engineers to focus more on design judgment and complex problem analysis. Additionally, the platform uses machine learning to predict manufacturing performance for carbon fiber structures and organizes past project insights via engineering knowledge graphs to support subsequent decisions.

Notably, Rescale claims the platform can boost professional engineering productivity by up to 3x, though public materials do not detail the measurement methodology, sample scope, or independent verification methods for this figure. Surrogate models do not fully replace traditional CAE; their predictive capability depends on training data and applicability ranges, meaning critical designs still require high-precision simulation confirmation. This case demonstrates the practical application value of AI in industrial R&D, extending beyond "code generation" to "physical simulation acceleration."

Quellen:TechOrange 科技報橘

Forschung

Apple's Inventory Surges 154% as CEO Cites 'Unprecedented' Memory Costs

Thema · 苹果内存成本冲击Wesentlich
2026-11-03 07:04 GMT+8

U.S. technology company Apple disclosed in its latest fiscal fourth-quarter results that its inventory levels rose significantly, growing from $11.09 billion in Q3 to $14.54 billion in Q4, a 31.1% increase quarter-over-quarter. Compared to $5.72 billion in the same period of 2025, inventory surged 154.3% year-over-year, reflecting the company's stockpiling strategy amid tight component supplies.

Apple CEO John Ternus stated during the earnings call that the current situation with memory pricing is "unprecedented." He noted that while carrying inventory and lower costs on certain non-memory components partly offset higher memory costs seen since June, the company expects even higher memory costs in the December quarter. Consequently, Apple lowered its gross margin guidance for the next quarter to between 45% and 46%, down from 47.9% in the September quarter.

Ternus also revealed that Apple's manufacturing purchase obligations increased by $33 billion from June to the September quarter as the company works to improve supply. Chief Operating Officer Sabih Khan joined the discussion on memory issues, reiterating the increased purchase obligations and not ruling out future price increases on devices. Additionally, weak iPad sales were attributed to supply constraints, with revenue at $6.2 billion, below analyst expectations of $6.7 billion and down 10% year-over-year.

Quellen:TechCrunch

Forschung

Anthropic updates Claude Code prompt to block video game clones

Thema · AnthropicWesentlich
2026-11-03 02:56 GMT+8

US-based AI company Anthropic has updated the system prompt for its coding tool, Claude Code, to explicitly prevent it from being used to recreate specific existing video games. The change was confirmed by Anthropic engineer Addy Osmani on November 2, 2026, aiming to curb growing issues with AI-generated game infringement.

Osmani stated that two key sentences were added to the prompt for the Opus 5.5 model: “Do build original games, games in a genre, and games from the user’s own designs and assets. Don’t build a recreation of a specific existing game, even if it is named or called original.” This follows reports from Tom's Hardware and screenshots shared by users on social media highlighting the new restrictions.

While direct cloning is blocked, Anthropic emphasized that users can still create original games, work within specific genres, or use their own assets. Osmani noted that this aligns with the company's broader usage policy against intellectual property infringement. He indicated that Anthropic would continue to iterate on these guidelines to balance creative freedom with legal compliance, inviting users to report false positives where the AI refused legitimate requests.

Quellen:The Verge

Forschung

Nasdaq 100 Enters Technical Correction, Down 10% From October Peak

Wesentlich
2026-11-03 06:09 GMT+8

As of the close on November 2, 2026, the tech-heavy Nasdaq 100 Index has fallen 10% from its peak in October, officially entering a technical correction. The S&P 500 Index declined approximately 4% over the same period.

The Philadelphia Stock Exchange Semiconductor Index (SOX) dropped 3.4% on Monday, adding to Friday's 7.2% plunge. The chip-stock gauge failed to rise above its 200-day moving average on Monday after closing below that level for the first time since June 2025.

"Lower oil should be a tailwind for tech, but the fact that stocks could not hold the open even with crude plunging tells shows how little conviction is left even after last week’s carnage," said Dave Mazza, chief executive officer at Roundhill Financial.

Quellen:econbrowser.com

Forschung

Vercel AI Gateway Adds Support for Meituan's LongCat 2.5 Model

Kurzfassung
Verifiziert 2026-11-03 06:44 GMT+8

Vercel has announced that its AI Gateway now supports the LongCat 2.5 model from Meituan. This model combines native visual understanding with coding and multi-step tool use capabilities, designed to assist agents working across complex codebases and full-project builds.

According to Vercel, LongCat 2.5 features improvements in visual programming, creative design, and 3D tasks, alongside longer agent workflows that carry engineering work from planning through execution. The model supports text and image inputs with a 1 million-token context window for handling code, documents, and extended task histories.

Developers can access the model by using meituan/longcat-2.5 as the model name. It is compatible with coding agents such as Claude Code, Codex, and Cursor, requiring the latest Vercel CLI installation and running npx vercel ai-gateway setup for configuration.

Vercel AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations. The gateway reflects provider pricing with no markup and does not charge a platform fee on inference, including for Bring Your Own Key (BYOK) requests.

Quellen:Vercel

Forschung

NVIDIA Encodes Kubernetes Adaptation Decisions for Vera Rubin NVL72 into Reproducible Recipes, Validating Cross-Node NCCL Performance for the First Time

Wesentlich
2026-11-03 02:00 GMT+8

Configuring Kubernetes for new GPU hardware is a process prone to errors. Drivers, kernels, container runtimes, networking, operators, and schedulers must work in concert, with settings tuned specifically to the hardware and workload. Even minor mismatches can prevent workloads from running or limit their performance. In a technical blog post dated November 2, 2026, NVIDIA cloud architect Mark Chmarny and software engineer Yuan Chen detailed how they completed this adaptation on early Vera Rubin NVL72 hardware and solidified their experience into versioned recipes within the NVIDIA AI Cluster Runtime (AICR).

The upgrade from GB200 NVL72 to Vera Rubin NVL72 involves more than fine-tuning; user-visible changes are significant. The GPU driver requires a new branch: the validated R615.23 is an open-source kernel module driver pre-installed on the host, while the official platform support target is R620. On the network side, ConnectX-7 NICs and BlueField-3 DPUs are replaced by ConnectX-9 SuperNICs and BlueField-4 DPUs, with DOCA 3.3.0 being the first GA release supporting ConnectX-9. The device topology also differs: the Vera Rubin NVL72 compute tray exposes eight groups of ConnectX-9 PCI devices totaling 32 functions, whereas the GB200 has only four ConnectX-7 adapters. Multi-node NVLink workloads still rely on IMEX, but DRA-managed IMEX is mutually exclusive with the host's static nvidia-imex service, which must be disabled.

These changes are scattered across various layers, from low-level bindings to the orchestration layer. AICR addresses this by encoding all findings into recipes: host-managed mode drivers (GPU Operator sets driver.enabled: false, while the container toolkit remains operator-managed), DRA drivers enabling ComputeDomains, a minimum Kubernetes version of 1.34.1 (1.34 provides the GA resource.k8s.io/v1 DRA API, and the .1 patch is required because the RKE2 CDI/NRI toolkit path needs containerd 2.1), Ubuntu 26.04 paired with a Vera-optimized arm64 64k-page kernel, and raising the NFD topology updater's default 60Mi memory limit—otherwise, it triggers OOM kills on multi-TB NVL72 worker nodes. The recipe also introduces DraNet for allocating ConnectX-9 devices to Pods and employs NodeWright to enforce an exclusive RDMA network namespace mode, ensuring each Pod sees only its own Host Channel Adapter.

The validation process executes in four steps on a live Vera Rubin NVL72 reference cluster. Step one generates the recipe, offline parsing 14 components and 5 overlay layers. Step two renders the deployment bundle, producing 61 files including ordered Helm charts and deployment scripts. Step three snapshots the cluster and runs validation: the deployment phase passes all four checks; the conformance phase passes eight, covering GPU Operator health, DRA/ComputeDomain support, KAI gang scheduling, Pod autoscaling, and secure accelerator access. Step four is the cross-node ComputeDomain test, verifying IMEX channel injection and NVLink fabric connectivity. The performance phase runs a cross-node NCCL all-reduce, completing a 16 GiB message all-reduce over the NVLink fabric via IMEX ComputeDomain, achieving a bus bandwidth of 744.71 GB/s, meeting the recipe's floor. The entire sequence (recipe, bundle, snapshot, deployment, and conformance validation) completes in about five minutes, with the performance phase adding the runtime of the NCCL job itself.

Quellen:NVIDIA Technical Blog

Forschung

Cisco DevNet Designer: HTML Prototypes Replace Figma as the Sole Source of Truth for AI Design Handoff

Thema · FigmaWesentlich
2026-11-03 01:23 GMT+8

The PubHub internal publishing tool at Cisco DevNet underwent a time-constrained redesign in 2026. Instead of following the traditional process of designing first and then implementing, the team had an AI coding agent generate end-to-end interfaces directly from user stories and backend APIs. The result was usable: users could complete the main flows, and the interface connected to the backend. However, designer Alejandro Marrufo and Head of UX Casey Tong wrote in a blog post that "usable" does not equal "finished"—the interface did not follow Cisco's visual language, some AI-generated components were unintuitive, and multiple workflows needed cleanup.

This raised a practical question: how can we leverage AI to accelerate design modifications and integrate them into the final product workflow? Their initial attempt involved converting existing HTML pages back into Figma, allowing designers to continue refining in a familiar tool, and then using a Figma MCP connection to sync changes back to code via the coding agent. The idea was appealing—keeping design work within Figma while AI handled the transfer to implementation.

In practice, three issues emerged. First, details were lost in transit: the round trip between Figma and code missed interactions, complete page and button flows, and animations—the very elements that make an experience feel "finished." Second, updates were slower than expected: modifying the HTML prototype via Figma was slower than the project timeline allowed; some pages even had to be rebuilt in Figma and reconnected to flows before any refinement could begin. Third, the source of truth became blurred: with Figma files and HTML prototypes evolving in parallel, it was difficult for designers and other team members to determine which version was current.

Figma is not without merit. It remains a superior tool for maintaining asset libraries and handling fine-grained UI elements. However, under tight deadlines, it is not the fastest place to refine an interface that is already running in code.

So they tried another path: letting AI coding tools inspect the existing prototype and modify its HTML directly. Design work remained within the existing interface, eliminating reconstruction steps. The HTML prototype became the single source of truth. Designers could make staged changes, review each iteration, and provide feedback to the AI along the way. Design work no longer required rebuilding the interface in a design tool first; instead, it targeted a running implementation directly, including interactions and flows that Figma struggled to capture.

A practical constraint quickly surfaced: token limits. Midway through the project, the designer ran out of tokens, switching from Cursor to Claude, then to Codex, resuming once limits reset. The work survived these switches because the HTML artifact—not any specific tool—held the state of the design. They recommend that teams adopting this approach plan for such constraints in advance.

With this approach, designers could edit and deliver working HTML pages daily, producing new iterations. Under the old Figma-first process, a single page took days. This speed created a compounding effect across the team: faster design output meant faster engineering testing, which in turn meant faster feedback for the next design iteration.

Quellen:Cisco Blogs

Forschung

NVIDIA RTX 5080 Price Surges Over 30% in Two Weeks; 24GB Variant Delayed

Thema · RTX 5080Wesentlich
2026-11-03 01:52 GMT+8

According to German tech outlet heise online, the price of NVIDIA's second-fastest high-end gaming graphics card, the RTX 5080, has risen by more than 30 percent in the last two weeks, making it difficult to find below 2,000 euros. This surge was partly fueled by earlier rumors that the flagship RTX 5090 might be discontinued. Although NVIDIA Vice President Jeff Fisher later denied these claims, market anxiety had already driven up demand for the secondary product. Meanwhile, the price of the RTX 5090 itself has climbed to around 8,000 euros.

More critically, the planned 24GB VRAM version of the RTX 5080, intended as a potential alternative or supplement to the RTX 5090, appears to be delayed. According to Taiwan-based hardware news site Benchlife, graphics card manufacturers have not yet received the complete design documents from NVIDIA for this new version, which were originally scheduled for October. Additionally, existing stocks of the GB203 GPU chips at manufacturing bases in China are sold out, with only enough supply to last one to two months. This casts significant doubt on the original plan to launch the new RTX 5080 during the CES exhibition in January 2027.

The root cause of this shortage is the ongoing artificial intelligence boom. NVIDIA uses the same GPU chips for its workstation graphics cards (such as the RTX Pro series), which feature larger memory capacities suitable for running local AI models and offer higher profit margins. For instance, the GB202 chip is used in both the RTX 5090 and professional cards like the RTX Pro 6000. With GDDR7 memory also in short supply, NVIDIA prioritizes the higher-margin professional workstation market, thereby squeezing the production capacity and pricing space for consumer gaming graphics cards.

Quellen:heise online

Forschung

Cisco Chief Engineer Julio Gomez: The Bottleneck in LLM Security Discovery Is Trust, Not Model Capability

Wesentlich
2026-11-03 00:30 GMT+8

Large models can confidently tell you there are vulnerabilities in your code, but you have no way of knowing which claims were actually checked, whether you will get the same answer next week, or why a specific finding exists six months from now when an auditor asks. Cisco Chief Engineer Julio Gomez argues that this gap represents the entire distance between LLM security tools moving from demonstrations to actual security programs.

He implemented the Foundry security governance specification released by Cisco—a design document rather than a product—and applied it to real code: the complete v1 API of Langflow 1.7.3 (289 function-level units across 25 modules), MLflow 2.9.0 (830 units), and OWASP's PyGoat training application (180 units). He set himself a hard rule: no strawman baseline. The control group used the same model, the same code, and a fair, disciplined, equally demanding prompt from a senior application security expert requiring structured findings and precise line-number citations.

The results were as follows: Foundry presented 32 findings for humans, while the baseline presented 15; both had 100% citation accuracy; but the baseline marked 14 out of its 15 findings as 'confirmed' with no mechanism enforcing that judgment, whereas Foundry's 28 confirmed and 4 pending-review items were delineated by mechanical checks. Gomez summarizes this by saying: The model was never the bottleneck; trust was.

There are two core design decisions in Foundry. First, detection is inclusive, while decision-making is strict. The Detector is instructed to prioritize recall—bringing forward anything plausible—all discipline being placed downstream. This allows the same model to be thorough and trustworthy simultaneously, rather than forcing a choice between the two. Second, the evidence gate is not a second opinion. It never asks the model 'are you sure?'. Instead, it takes every citation back to verify against the bytes on disk: the file must exist in the target, the line number must be within range, and the quoted text must actually appear there. Only findings whose citations survive are labeled 'confirmed'; others are downgraded to 'pending review' or rejected outright. A model cannot argue its way past a line number that does not exist.

Gomez illustrates why this separation matters through repeated runs. He independently ran the full pipeline twice; the 122 records stored collapsed into 67 distinct fingerprints, with 56 recurring across both runs and landing on existing identities—roughly 90% overlap, despite the inherent non-determinism of model output. This means the next CI run can calculate deltas (new, recurring, gone) rather than forcing the team to re-read a wall of differently worded text. After the second run, teams usually stop running it again; once stable identities are established, they continue.

On the broader scope of MLflow, governance held true: 830 units, 41 batch calls, and 4.8 million tokens completed within the declared 6 million cap. Findings landed in the same area as MLflow's historical CVEs (path traversal in the artifact storage layer). However, only 13 were labeled confirmed, while 110 were leads pending review—the gate became more conservative on large-scale unfamiliar code, which is precisely what the layered approach is doing right.

Quellen:Cisco Blogs

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

However, boundaries must be clearly understood. The recipe is in Preview status and does not carry production support or lifecycle qualification. Validation was self-reported by the vendor on NVIDIA reference hardware, with no independent reproduction. Inference serving performance is explicitly deferred: vLLM merged Rubin (sm_107a) support code in v0.27.0, but the SM107 kernel path is gated at build time on CUDA ≥ 13.4, while vLLM’s published images up to v0.28.0 are still built against CUDA 13.0.3. Thus, the latest image carries dispatch code but lacks compiled kernels. ai-dynamo’s latest release, v1.4.2, pins vllm-openai v0.26.0, which predates this code. An available inference performance floor requires a published runtime image built on CUDA 13.4 and verified on real hardware.

The validation process itself drove upstream fixes in reverse. When running on early silicon, AICR discovered an unmapped Rubin architecture value in go-nvlib that caused the DRA driver’s kubelet plugin to crash-loop. Because architecture support lands first in go-nvml, flows into go-nvlib, and then into the GPU Operator, Kubernetes Device Plugin, MIG Manager, and DRA plugins, the fix propagated consistently down the stack. From the first crash report to a validated build on Vera Rubin hardware took about three days, finally shipping with the official DRA driver v0.5.0 release. Validation also revealed that host-managed IMEX daemons and DRA-managed ComputeDomains cannot currently share a node, feedback that shaped the host-managed IMEX option added in DRA driver v0.5.0.

For platform teams, the practical value of this record lies not in the number 744.71 GB/s itself—it is a vendor self-test result on reference hardware—but in how it demonstrates that the complete decision list for new hardware adaptation is solidified into reproducible, verifiable recipes. This allows a second cluster to repeat the same results the next day without having to re-navigate pitfalls like NFD memory limits or IMEX exclusivity conflicts.

The resulting workflow is iterative: the product team defines user stories and backend APIs; an AI coding agent builds an end-to-end interface based on these; designers review the working HTML pages and flows, refining the prototype with AI coding tools; designers hand over the updated HTML artifacts to developers, along with the design system, end-to-end flows for different personas, and a changelog describing the changes; developers pass these to a coding agent, which applies the design changes to the project; the team tests and reviews the results, often live, and repeats the cycle.

The HTML prototype is no longer just a mockup; it has become the handoff document itself. Engineers can use it to accurately reproduce the design, and this approach proved surprisingly popular. Several factors made this possible: an embedded design system (colors, fonts, components, and spacing are placed alongside the page, allowing engineers to reference rules rather than infer them from screenshots); a persona switcher (controls within the prototype allow switching between different user roles, giving engineers clear visibility into each flow and its variants); and a changelog (recording changes per iteration, allowing engineers to apply updates incrementally rather than comparing entire pages).

Regarding syncing design and code, they initially prompted the coding agent with issue lists to fix problems one by one, which was slow and manual. Later, they required designers to include a changelog with each new version. Engineers now hand over both the changelog and the updated HTML to the agent, which automatically applies changes across the project. The handoff shifted from one-off prompts to largely automatic update steps.

Their biggest takeaway was not that more code was written using AI, but that the feedback loop between design and engineering was brought closer together, blurring the lines between the two roles. Reviews became more frequent, with designers sometimes editing in real-time during meetings. Engineers began testing earlier, even performing minor design iterations themselves. The whole team could spend more time on experience and vision, and less on implementation logistics.

For teams wishing to try this method, they listed several practical recommendations: start with a working build, having the AI agent generate an end-to-end interface from user stories and API specs, then design against it rather than starting from a blank canvas; maintain a single source of truth, making the HTML prototype the canonical design and using Figma only where it excels (asset libraries, fine details), avoiding parallel copies of every screen; embed handoff into the prototype, including the design system, changelogs for each iteration, and methods to view all persona or flow variations; explicitly specify what the agent cannot infer, documenting secondary interactions and edge cases so the coding agent doesn't invent them; plan for token limits, expecting to switch AI tools mid-project since the HTML artifact holds the design state, allowing work to continue; shorten review cycles by reviewing more frequently and being open to real-time edits with engineers present; keep the designer's judgment central—AI can quickly produce a functional first draft, but it does not know your design language or which interactions feel right.

For them, this experiment yielded a practical pattern: let AI create the first usable version, let designers refine a functional prototype that also serves as the handoff document, and let the coding agent carry those refinements into the project. This provided a faster path—from an interface that works to one that feels like their own product.

Gomez explicitly states he is not claiming Foundry is smarter or cheaper. With the same model, Foundry surfaces more findings because it performs exhaustive scanning rather than letting a single agent turn curate the results. Token costs are approximately ten times that of a single call, buying coverage, the mechanical gate, and deduplication. 'Confirmed' means verified and worth a reviewer's time, not exploited—proving exploitability requires a live testbed, which is a separate role not implemented in the specification.

His conclusion is that the value of the governance layer only becomes apparent when findings need to integrate into a process: an issue tracker that cannot be drowned in duplicates, reviewers who need to know what was checked, and pipelines that repeat as code changes. For a one-time check by an engineer looking at a small piece of familiar code, a standard agent is cheaper and more practical; just use it directly.

This article is the second part of the Security Loop series, published on November 2, 2026. All run logs, provenance trails, rule gaps, and scoring comparisons are publicly available at github.com/juliogomez/foundry.