☕️ Agentic AI breaks old data architectures

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

☁️ AWS hit by routing failure after Azure

🏗️ Agentic AI breaks old data architectures

🇨🇳 Kimi K3 weights open with conditions

🟣 Microsoft ships.NET 11 Preview 6

🐛 Cloudflare open-sources a privacy-protocol debugger

Plus: 🎁 13 other news you might like, 🧰 6 tools, and 📚 5 papers.

☁️ AWS hit by routing failure after Azure LINK

AWS suffered a routing failure in its us-west-2 region on July 24, cutting internet connectivity and Direct Connect access, a day after Azure similarly isolated a West US datacenter despite workloads continuing to run.

The outage began at 10:55 UTC and was traced to networking hardware serving traffic into and out of us-west-2, with Direct Connect users routing through the EqSe2 exchange in Seattle hit by an extended disruption.

An initial mitigation restored most connectivity, but route reconvergence triggered a second drop between 11:47 and 11:59 UTC before service was fully recovered.

🏗️ Agentic AI breaks old data architectures LINK

Agentic AI systems that perceive, decide, and act are breaking traditional data architectures, prompting the AI gateway pattern: a single control plane isolating fast-changing guardrails, model routing, agent identity, and audit from stable enterprise platforms behind it.

The pattern exists because agentic systems violate API gateway assumptions (determinism, schema-level failures, and client-specified actions) so a validly formatted, authenticated request can still issue a wrong refund or leak data through prompt injection or context poisoning.

The gateway centralizes per-action zero-trust policy per NIST SP 800-207, semantic logging of Request→Decision→Action, and policy-as-config declared in version-controlled PRs, with kgateway's Agent Gateway, Portkey, and LiteLLM already implementing subsets of these responsibilities.

🇨🇳 Kimi K3 weights open with conditions LINK

Moonshot has released the weights for Kimi K3, its 2.8-trillion-parameter multimodal model, letting developers download, fine-tune, and run it locally for free after initial demand overwhelmed the company's systems.

K3 targets long-horizon coding, knowledge work, and reasoning, with Moonshot claiming it performs competitively against Fable 5 while substantially outperforming Opus-4.8, GPT-5.6 Sol, and GPT-5.5 across various benchmarks.

The open release should benefit cloud providers that specialize in hosting third-party models, though Moonshot has paused new subscriptions after a sixfold jump in daily sales, offering a waitlist instead.

🟣 Microsoft ships.NET 11 Preview 6 LINK

Microsoft has released.NET 11 Preview 6, delivering updates across the C# language, ASP.NET Core,.NET MAUI, and Entity Framework Core, with new features spanning extension indexers, union types, and cross-origin request protection.

C# extension members now support indexers alongside methods and properties, while union types ship compiler support so declarations compile without boilerplate and System.Text.Json serializes the active case, both still requiring the preview language version.

ASP.NET Core apps using WebApplication.CreateBuilder automatically reject unsafe cross-origin requests via Sec-Fetch-Site and Origin headers, Minimal API validation gains async validators, and generated OpenAPI documents now target version 3.2 by default.

🐛 Cloudflare open-sources a privacy-protocol debugger LINK

Cloudflare has released pvcli, a command-line debugger for Oblivious HTTP and eventually MASQUE traffic under an MIT license, targeting the split-trust privacy infrastructure that powers services like Apple's iCloud Private Relay.

The tool tackles a diagnosis problem baked into these protocols: requests traverse multiple independently operated systems where no party sees the full path, so pvcli lets teams test proxied traffic against their own infrastructure or a sandbox at ohttp.info.

Its command syntax is modeled closely on curl, deliberately built so AI agents can construct and debug requests autonomously, and Cloudflare tested it with an unnamed set of beta customers before this launch.

Other news you might like

• Tailwind CSS vs. StyleX: A real migration with 20 componentsLINK

• Building Quill: A Slack AI Blog Agent with AstroLINK

• The harness is all you need (mostly)LINK

• Why You Probably Don't Need A Supervisor MCPLINK

• I don't trust Wi-Fi routers anymore, so I built my ownLINK

• 12 Startups in 12 MonthsLINK

• I tested a new Debian Linux loaded with local AI - see if it's right for youLINK

• The true test dataset for a generalised taskLINK

• Gamed Benchmarks, Context Anxiety, and LoRA's LimitsLINK

• MSE loss does not generate superpositionLINK

• Java News Roundup: Simple JSON API, JEPs for JDK 28, Oracle CPUs, Embabel 1.0, Azul Payara, HelidonLINK

• How NVIDIA Builds Open Models for the Age of AILINK

🧰 Trending tools

AnySearch: a search API built for AI agents that pulls filtered, de-duplicated, structured results from trusted sources in parallel, improving reliability.LINK

Sim: a workspace for building and deploying AI agents visually or with code, connecting to 1,000+ integrations and every major LLM provider.LINK

Zro: routes coding requests to open-source models like MiniMax M3, GLM-5.2, and Kimi K2.7 across regions without retaining any data.LINK

CartAI: lets developers automate checkout on any merchant site through one API, managing payments, shipping, and order tracking without requiring merchant integration.LINK

QApilot's CoWork: converts existing test cases into automated mobile tests using AI planning, human approval checkpoints, and real-device execution across iOS, Android, and Flutter.LINK

Loomal: lets AI agents pay for API access or store purchases via USDC, settling on Base in seconds with no fees.LINK

📚 Trending papers & reports

Computer-use agents that read a program's actual files and data instead of just screenshots boost task success from 20.6% to 26.9%, at roughly 9 times lower cost per task.LINK

System log analysis gets a purpose-built upgrade for spotting outages, cutting to the root cause and identifying fault types, beating standard language models by up to 9.83% in detection accuracy, 18.28% in root cause ranking, and 20.88% in fault classification.LINK

Quantum program testing gets a noise-cleanup tool that spots real software bugs on noisy quantum hardware with up to 94.90% precision, making today's error-prone quantum computers usable for reliable software testing.LINK

Automated bug hunting for blockchain smart contracts now lets AI write test checks nearly as effective as human experts, catching costly flaws before contracts go live and become unfixable, with Gemini Pro 1.5 scoring 25.99% versus humans' 31.75%.LINK

Software impact prediction combines a map of how code pieces connect with deep learning trained on code text, more accurately predicting which parts of a program break when a change is made, without needing execution logs or years of change history.LINK

See you tomorrow for a new dose of ☕️ Devshot!