Microsoft open-sources AI unit test agent with 92% completion rate across six languages

<p>Microsoft’s .NET team has released <a href="https://github.com/dotnet/skills" target="_blank" rel="noopener nofollow">code-testing-generator</a>, an open-source agent that writes unit tests across more than a dozen languages. It’s available now via GitHub Copilot CLI, with VS Code support in preview.</p>
<h2>What makes it different from a plain Copilot prompt</h2>
<p>Asking Copilot to “generate unit tests” works, sort of. The new agent takes a different approach: before writing a single test, it scans the repository to detect the language, test framework, existing conventions, and build commands. Then it picks one of three execution paths based on scope:</p>
<ul>
<li><strong>Direct</strong> — for small, well-defined tasks</li>
<li><strong>Single-pass</strong> — for medium-scope work</li>
<li><strong>Iterative</strong> — for larger or more ambiguous tasks</li>
</ul>
<p>From there it writes tests incrementally, starting with simpler code paths and working toward complex dependencies. As it goes, it fixes compilation errors, avoids flaky patterns like real URLs or timing-sensitive assertions, and runs the full test suite to confirm everything passes.</p>
<p>The last step is lightweight mutation testing — deliberately breaking the code under test to verify the tests actually catch bugs, not just execute without failing.</p>
<h2>Benchmark results</h2>
<p>Microsoft ran the agent against stock GitHub Copilot across 152 real-repository tasks:</p>
<table>
<thead>
<tr>
<th></th>
<th>code-testing-generator</th>
<th>Stock Copilot</th>
</tr>
</thead>
<tbody>
<tr>
<td>Task completion</td>
<td>92.1%</td>
<td>78.9%</td>
</tr>
<tr>
<td>Failure reduction</td>
<td>63% fewer</td>
<td>—</td>
</tr>
<tr>
<td>Speed</td>
<td>5.5% faster</td>
<td>—</td>
</tr>
</tbody>
</table>
<p>The gap was sharpest on vague prompts (88.8% vs. 66.3%) and on diff-specific prompts, where the agent went 15-for-15 while stock Copilot went 0-for-15. Coverage and test count were roughly comparable — the agent produced slightly fewer tests at similar coverage, which suggests it’s not padding output.</p>
<p>The workflow held up across Claude Opus 4.8, GPT-5.5, and Claude Haiku 4.5, so the gains aren’t model-specific.</p>
<h2>Language support and limitations</h2>
<p>The agent supports .NET, Python, Go, TypeScript, Java, Rust, and more. It’s explicitly scoped to unit tests only — integration, end-to-end, browser, and performance tests are out of scope for now.</p>
<p>The repository is at <a href="https://github.com/dotnet/skills" target="_blank" rel="noopener nofollow">dotnet/skills</a> on GitHub.</p>


Microsoft's .NET team has open-sourced code-testing-generator, an AI agent that writes and validates unit tests across multiple languages including .NET, Python, Go, TypeScript, Java, and Rust. Unlike simple model prompting, it uses a 'research-plan-implement' (RPI) pipeline that first reads the repository to understand language, test framework, conventions, and build commands. It then selects one of three execution paths (direct, single-pass, or iterative) based on task complexity. Benchmarked against 152 internal tasks, it achieved a 92.1% completion rate versus 78.9% for plain GitHub Copilot, with 63% fewer failures overall. The agent is scoped to unit tests only and is available in the dotnet/skills GitHub repository, working with GitHub Copilot CLI and VS Code.

Node.js 26.7.0 ships Perfetto tracing and updated root certificates

Node.js 26.7.0 (Current) has been released with several notable changes: Perfetto tracing support has been added for improved diagnostics, the crypto module now supports loading private keys through STORE loaders and updates root certificates to NSS 3.125, Symbol.dispose is implemented in ModuleHooks for explicit resource management, and the test runner gains a --test-coverage-include-all flag. The release also includes numerous bug fixes across crypto, FFI, SQLite, streams, HTTP/2, QUIC, and TLS, along with dependency updates including npm 11.19.0, SQLite 3.53.4, and various V8 backports.

Google ships official arm64 Linux builds of Chrome

Google has shipped official arm64 Linux builds of Chrome as of July 30, 2026. The builds are available as .deb and .rpm packages via the Chrome download page and Google's apt/RPM repositories. For teams running browser automation, CI/CD testing, or synthetic monitoring at scale, this enables running Chrome on more energy-efficient arm64 hardware (e.g., AWS Graviton), which uses up to 60% less energy and costs up to 20% less than comparable x86 instances. Chrome for Testing arm64 builds are also confirmed to be in progress. The author, who filed the original feature request in October 2024, notes this is a significant win for operational efficiency and sustainability at scale.

Bun adds test timings and intelligent CI shard splitting

An upcoming Bun release introduces `bun test --timings` to record per-file test durations, and `bun test --parallel --shard=i/N` which leverages those timings to distribute CI shards using longest-processing-time-first scheduling, optimizing large test suite execution across parallel runners.

Next.js 16.3 ships instant navigations, 90% less dev RAM and 5.5x faster CI builds

<p>Next.js 16.3 is out, and it’s a substantial release. The headline numbers are real: up to 90% less memory in the dev server, faster builds, and measurable infrastructure savings at scale. Here’s what’s actually in it.</p>
<h2>Dev server memory</h2>
<p>The notorious <code>FATAL ERROR: JavaScript heap out of memory</code> crash has been a thorn in the side of anyone running complex Next.js apps locally. The culprit was V8 heap exhaustion under Turbopack. In 16.3, Turbopack gets disk caching and memory eviction enabled by default, dropping RAM usage from roughly 4.6GB to around 840MB for a 50-route app. That’s not a rounding error.</p>
<p>Build times improve too. Filesystem caching gives a 2.3x compilation speedup in CI, and there’s an experimental Rust-based React Compiler baked into Turbopack that cuts cold build times by 34% and warm builds by 46%.</p>
<h2>TypeScript 7</h2>
<p>TypeScript 7 support lands in this release, bringing roughly 10x faster type checking. If you’ve been sitting on a large codebase watching <code>tsc</code> grind away, this is the one to notice.</p>
<h2>Instant Navigations</h2>
<p>This is the feature getting the most attention. Instant Navigations is an opt-in mode that makes server-driven apps feel like SPAs, without requiring client-side data fetching. It works through a combination of Partial Prefetching, improved ISR, and a new <code>cacheComponents</code> option that lets Server Component pages either stream with loading shells or serve cached content immediately.</p>
<p>The Vercel team applied this to v0 using an AI agent loop: the agent wrote failing Playwright tests using a new <code>instant()</code> test helper, then fixed the code (mostly moving dynamic data fetches behind Suspense boundaries) until the tests passed. The result was 16 regression tests and instant navigations across v0’s homepage, chat, and settings pages. A <code>next-cache-components-optimizer</code> skill is available if you want to run the same process on your own app.</p>
<p>There’s also a Navigation Inspector panel in DevTools for debugging this, and the plan is to make Instant Navigations the default eventually.</p>
<h2>Runtime performance</h2>
<p>Native Node.js streams now replace web stream conversions for SSR, which gets you about 22% more request throughput. Not glamorous, but it adds up.</p>
<h2>Platform-level improvements on Vercel</h2>
<p>For teams running on Vercel, 16.3 brings some concrete infrastructure wins:</p>
<ul>
<li>45% fewer prefetch requests</li>
<li>17% fewer CDN requests and 24% fewer bytes transferred for static content</li>
<li>~2x faster p99 route resolution for large sites</li>
<li>Immutable static assets reused across deployments, cutting re-upload time by up to 30%</li>
<li>A revamped route metadata layer using JSONL shards, resulting in ~10x fewer cache misses</li>
</ul>
<p>The routing metadata improvements apply to any framework using Vercel’s Build Output API, not just Next.js.</p>
<p>PPR observability is also launching with this release, so you can see which requests are serving static vs. dynamic content.</p>
<h2>Agent-native tooling</h2>
<p>There’s a deliberate push here to make Next.js easier to work with through AI coding agents. That includes officially maintained versioned docs (so agents don’t hallucinate outdated APIs), built-in skill sets for common migrations, and an ‘Instant Insights’ feature that detects caching errors and generates prompts for agents to fix them.</p>
<p>The framing from the team is that when Instant Navigations is on, the agent is essentially forced to make things fast — it gets presented with choices like adding a <code>&lt;Suspense fallback&gt;</code> rather than blocking navigation on a server fetch.</p>
<h2>Other things worth noting</h2>
<ul>
<li>Custom error boundaries with server retry</li>
<li>Bundled prefetch requests</li>
<li>Experimental network resilience with offline support</li>
<li>Next.js has now passed one billion downloads</li>
</ul>
<p>Andrew Clark (acdlite) noted that some of what shipped here traces back to the first Suspense prototype in 2017. That’s a long arc. The release had 90 contributors.</p>
<p>Upgrade path is reportedly smooth — the team says agents handle the migration well, particularly for enabling Instant Navigations.</p>


Next.js 16.3 has been released with significant performance improvements, including up to 90% reduction in memory usage during development, faster builds, faster type checking, faster rendering, and improved tooling. The announcement is light on technical details beyond these headline metrics.

Node.js 26.6.0 adds FFI event loop API and test runner improvements

Node.js 26.6.0 has been released with several notable changes including a new FFI `getCurrentEventLoop` API, two test_runner improvements (`context.log()` method and `test:log` event, plus `entryFile` reporting in TestStream events), and a large number of bug fixes and performance improvements across streams, HTTP/2, QUIC, crypto, and other subsystems. Dependency updates include npm 11.18.0, libffi 3.7.1, ngtcp2 1.24.0, and zlib 1.3.2.1.

Tricentis acquires Tabnine to add AI knowledge graph to testing platform

<p>Tricentis has acquired Tabnine, the enterprise-focused AI coding platform, to integrate Tabnine’s Enterprise Context Engine into its Agentic Quality Engineering Platform.</p>
<p>The core of what Tricentis is buying is a knowledge graph. Tabnine’s Enterprise Context Engine builds a hybrid graph-plus-vector model of an organization’s systems by pulling entities, relationships, dependencies, and architectural patterns from code repositories, documentation, tickets, APIs, and infrastructure metadata. The idea is that AI testing agents need to actually understand what they’re working with before they start acting — knowing the architecture, the dependencies, the organizational context — rather than operating blind.</p>
<p>Practically, this adds a few things to the Tricentis platform: real-time organizational intelligence, dependency and impact analysis, automated governance, and shared memory across multi-agent workflows. That last one matters more than it might sound. When multiple agents are collaborating on a task, shared context means they’re not each starting from scratch or working at cross-purposes.</p>
<p>The reported numbers are notable: organizations using the technology have seen up to 2x improvement in AI accuracy, 80% reduction in token consumption, and 50% faster resolution on complex tasks. Token reduction in particular is a meaningful cost and performance metric for anyone running AI agents at scale.</p>
<p>The acquisition also addresses a problem Tricentis has been tracking. A recent Tricentis survey found that 60% of organizations already regularly ship untested code to production. AI-generated code is accelerating output faster than QA processes can keep up, and the argument here is that smarter, context-aware testing agents are one way to close that gap.</p>


Tricentis has acquired Tabnine, the enterprise AI coding platform, primarily to integrate Tabnine's Enterprise Context Engine into its Agentic Quality Engineering Platform. The Context Engine builds a structured knowledge graph of an organization's codebase by extracting entities, relationships, dependencies, and architectural patterns from repositories, docs, tickets, APIs, and infrastructure metadata. This gives Tricentis's AI testing agents a shared memory layer for real-time context sharing. The move is driven by a Tricentis survey finding that 60% of organizations regularly ship untested code to production — a gap widening as AI-generated code outpaces QA processes.

Pest 5 ships test impact analysis cutting Laravel Cloud's test suite from 3 minutes to 5 seconds

Laravel ships LSP support, AI agent tools, and scale-to-zero cloud at Laracon 2026

<h2>Overview</h2>
<p>Laracon US 2026 packed in a lot of announcements across the Laravel framework, tooling, and Cloud platform. Here’s a rundown of what shipped or was previewed.</p>
<hr>
<h2>Laravel LSP</h2>
<p>Laravel now has a first-party language server. Laravel LSP understands your application’s context and provides completions for config keys, route names, view paths, translation strings, middleware, Livewire components, and more — not just generic PHP syntax.</p>
<p>Official editor support covers VS Code, Sublime Text, Zed, Neovim, Cursor, and OpenCode. The Zed extension (v0.1.0) is already in the extension registry and handles binary downloads automatically. Worth noting: a community Zed extension already existed and takes a different approach (tree-sitter parsing rather than running PHP scripts), so you shouldn’t run both at once.</p>
<hr>
<h2>Blade formatting in Laravel Pint</h2>
<p>Pint 1.30.0 adds Blade template formatting behind a <code>--blade</code> flag. Under the hood it runs <code>.blade.php</code> files through Prettier using the <code>prettier-plugin-blade</code> plugin, so Node.js and a few npm packages are required. If they’re missing, Pint detects your package manager and offers to install them.</p>
<p>The formatter handles Alpine attributes, Livewire <code>wire:</code> bindings, Flux components, and Tailwind class sorting. A few file categories are excluded — Envoy tasks, mail views, and Boost guidelines — to avoid unintended output changes. The feature is fully opt-in, so existing projects aren’t affected.</p>
<p>The release also fixes <code>--diff</code> to include renamed files and corrects <code>--quiet</code> to write violations to stderr rather than silently dropping them. There’s also a new <code>in</code> key in <code>pint.json</code> for restricting which directories get scanned.</p>
<hr>
<h2>Laravel Doctor</h2>
<p><code>php artisan doctor</code> is a new first-party package (v0.1) that runs health checks against your application. It covers:</p>
<ul>
<li>Environment setup (APP_KEY, PHP version, required extensions)</li>
<li>Composer dependencies and <code>composer audit</code> for security vulnerabilities</li>
<li>Database connectivity</li>
<li>Cache, queue, and session driver configuration</li>
<li>Storage permissions</li>
<li>Security settings</li>
</ul>
<p>Checks return one of six statuses. Many issues can be auto-fixed with <code>--fix</code>. It’s environment-aware (local vs. production behavior differs), supports custom diagnostics registered by third-party packages, and outputs JSON or GitHub Actions-compatible formats for CI. There’s also an agent-optimized output mode for tools like Claude Code or Cursor.</p>
<p>Requires PHP 8.3 and Laravel 12 or 13.</p>
<hr>
<h2>Laravel Head</h2>
<p>Laravel Head is a new official package for managing everything inside an HTML document’s <code>&lt;head&gt;</code>. It provides a fluent API for:</p>
<ul>
<li>A five-layer metadata precedence system</li>
<li>Route-level metadata via <code>withHead()</code></li>
<li>Open Graph and Twitter card tags</li>
<li>JSON-LD structured data (articles, products, breadcrumbs, FAQs, etc.)</li>
<li>Resource hints (preload, prefetch, preconnect)</li>
<li>PWA metadata and favicons</li>
</ul>
<p>It works with Blade, Livewire, and Inertia.js including SSR. Requires PHP 8.3 and Laravel 13.17+.</p>
<hr>
<h2>Laravel AI SDK: human-in-the-loop approvals</h2>
<p>v0.10.0 of the Laravel AI SDK adds a human-in-the-loop API so agents can pause before running sensitive tools and wait for explicit approval. Tools opt in via the <code>Approvable</code> contract and <code>InteractsWithApprovals</code> trait. Approval logic can be conditional — for example, only requiring a human sign-off for refunds above $2,000.</p>
<p>When a tool is paused, the response exposes pending approvals with the tool call ID, name, arguments, and reason. Resuming requires passing a <code>Decisions</code> object with explicit approve or reject choices per call. The feature works across prompt, stream, queue, and broadcast modes.</p>
<p>Upgrading requires a schema migration to add an <code>approval_state</code> column and implementing <code>storeApprovalResults()</code> in any custom <code>ConversationStore</code> implementations.</p>
<p>The SDK also gained filesystem tools for agents, a <code>summarize()</code> string helper, multimodal embeddings, and OpenAI-compatible provider support.</p>
<hr>
<h2>Laravel Cloud updates</h2>
<p>Cloud shipped scale-to-zero Flex compute with cold starts under 500ms — MySQL included, which is the part that actually matters since databases are usually what prevent scale-to-zero from being practical. Managed queues now support autoscaling and FIFO. There’s also a new secrets manager, monorepo deployment support for Next.js and Nuxt, and Private Cloud updates including HIPAA compliance.</p>
<hr>
<h2>Other announcements</h2>
<ul>
<li><strong>Image manipulation API</strong> — a new built-in API for image processing</li>
<li><strong>CPX</strong> — an <code>npx</code>-equivalent for PHP packages</li>
<li><strong><code>artisan dev</code> command</strong> — development workflow improvements</li>
<li><strong>Inertia DevTools</strong> — a Chrome extension for debugging Inertia apps</li>
<li><strong>Refreshable locks and debounced jobs</strong> — concurrency and queue improvements</li>
<li><strong>Package skeleton template</strong> — updated scaffolding for new packages</li>
<li><strong>Pest v5</strong> — opened the conference as the first talk</li>
</ul>
<p>Tanner Linsley also mentioned that TanStack Octane support for TanStack Start is nearly complete, with adapters either done or in progress for their other libraries.</p>


Laracon US 2026 keynote introduced a wide range of Laravel updates. Key framework additions include Laravel LSP for multi-editor language server support (NeoVim, Zed, Sublime Text, Cursor, OpenCode), Blade file formatting in Pint 1.30, and a new `artisan doctor` health-check command covering environment config and security audits. Other framework additions include `artisan dev`, a head tag API, refreshable locks, debounced jobs, and a CPX tool similar to npx for PHP. On the AI SDK front, human-in-the-loop tool approval, filesystem tools for agents, multimodal embeddings, and OpenAI-compatible provider support were added. Laravel Cloud gains scale-to-zero Flex compute with sub-500ms cold starts (including MySQL), autoscaling managed queues with FIFO ordering, a built-in secrets manager, monorepo deployments for Next.js and Nuxt, and HIPAA compliance for Private Cloud.

Antithesis finds bugs in every Raft implementation it tested including HashiCorp and Aeron

Biome v2.5 ships 500 lint rules and cross-file CSS class linting

Biome v2.5 has been released, surpassing 500 lint rules with 73 nursery rules promoted to stable. Key highlights include two new cross-language lint rules (noUndeclaredClasses and noUnusedClasses) that leverage the module graph to detect CSS class mismatches across HTML/JSX/CSS files, a noRestrictedDependencies rule based on the e18e initiative, and GritQL plugin improvements including path-specific plugin application and code fix support. The release also adds a new delimiterSpacing formatter option for accessibility, a --watch mode for lint/format/check commands, go-to definition support in the language server, a linter preset option replacing the deprecated recommended flag, a biome upgrade command, a concise reporter for AI coding agents, and ~13% performance improvements. SVG file support, pnpm catalog resolution, and new sorting assist actions for package.json and other file types are also included.

RubyMine 2026.2 ships agentic debugging and native GitHub Copilot integration

RubyMine 2026.2 ships with several notable additions: agentic debugging lets AI agents autonomously run debug sessions and inspect runtime state using a new rubymine-debugger skill; native GitHub Copilot integration is now built in via a JetBrains-Microsoft partnership; AI code completion now supports OpenAI-compatible third-party model providers. The symbol-based code insight engine introduced experimentally in 2026.1 is now enabled by default, bringing more accurate navigation, Quick Documentation with type info, and better handling of partially resolved code. Other improvements include expanded RSpec 4 support with deprecation detection and quick-fixes, better RBS type hints, a more reliable 'Ignore non-project sources' debugger option, faster background gem environment updates, automatic Ruby interpreter detection, and optimized RuboCop background processing.

Stripe open-sources benchmark showing AI agents struggle with validation in real integrations

Stripe has open-sourced a benchmark suite testing whether AI agents can build real-world Stripe integrations end-to-end across backend, frontend, and browser-based checkout workflows. The benchmark uses 11 reproducible environments with full codebases, databases, and test API keys, evaluated via a Goose/MCP harness. Claude Opus 4.5 scored 92% on full-stack API tasks while GPT 5.2 reached 73% on structured tasks. Two key failure modes emerged: agents misinterpreting HTTP 400 error responses as success signals during SDK upgrades, and failing to recover from disrupted browser state during checkout flows. The core limitation identified is not code generation but validation reasoning, state management, and recovery in multi-step workflows — areas critical for financial systems where partial success is unacceptable.

GNOME builds TestFlight-style Test Center app for experimental features

<p>GNOME developers are prototyping a new app called <strong>Test Center</strong>, designed to make testing experimental features and development builds safer and more accessible, particularly on image-based systems like GNOME OS.</p>
<p>The project draws inspiration from Apple TestFlight and the old Mozilla Labs. The basic idea: let users try unstable or experimental software without risking their system.</p>
<p>For apps, developers can share a link to a CI build that installs as a tagged experimental Flatpak with automatic expiry. For system components, the approach uses systemd-sysext images as overlays rather than replacing anything outright, so removing them fully restores the previous state. No permanent changes, no cleanup headaches.</p>
<p>Test Center is part of the broader GNOME OS Developer Tool Suite and has been funded by Germany’s Prototype Fund since June 2026. The team is also working on easier CLI tool distribution for image-based systems, using Buildstream as part of the toolchain.</p>
<p>The project is still in the prototyping phase.</p>
<p>In other GNOME news this week: Glycin gained OpenEXR image support along with pixel density reading for JPEG, PNG, and TIFF formats, and GLib received improvements to network monitoring.</p>


GNOME developers are prototyping Test Center, an app inspired by Apple TestFlight and Mozilla Labs that lets users safely try experimental features and development builds. For apps, developers share CI build links that install as tagged experimental Flatpaks with automatic expiry. For system components, systemd-sysext overlay images are used so removal fully restores the previous state with no permanent changes. The project is part of the GNOME OS Developer Tool Suite, funded by Germany's Prototype Fund since June 2026. Also in GNOME news: Glycin gained OpenEXR support and pixel density reading for common image formats, and GLib received network monitoring improvements.

Angular 22 ships signals as stable with new forms API and SSR improvements

Cypress 16.0.0 deprecates bundled Electron browser in favor of Chrome for Testing

Cypress is deprecating the bundled Electron browser starting with Cypress 16.0.0, with full removal planned in a future major version. The decision stems from three core issues: Electron's Chromium version lags behind stable Chrome, Electron behaves differently from real browsers causing hard-to-debug test discrepancies, and security updates are delayed through the Electron release chain. Teams are advised to migrate to installed browsers like Chrome for Testing. The post provides a step-by-step migration guide covering how to install a supported browser, set a defaultBrowser in config, update CI pipelines, switch Docker images, and remove all Electron-specific code and environment variables. Chrome for Testing is recommended as the preferred test browser due to its pinned versioning, no auto-update behavior, and automation-friendly design.