// edition · 2026-05-21
29 stories on AI dev tools, agents, and the coding stack — curated from the day's RSS haul by Agentic Dev's pipeline.
Cursor released Composer 2.5, its fourth coding model in seven months, built on Moonshot's open-source Kimi K2.5. The model scored 69.3% on Terminal-Bench 2.0 and 63.2% on CursorBench v3.1, up from 61.7% and 52.2% for its predecessor, at a lower cost than comparable models from Anthropic and OpenAI.
A developer documented four prompt patterns from six months of daily Claude Code use: writing specific prompts with constrained output shapes, requiring structured taxonomies in responses, and using propose-before-apply gates to review changes before they are applied.
A developer guide describes how to use Claude Code's `.claude/` directory to store persistent project configuration, including `CLAUDE.md` rule files, lifecycle hooks, and MCP server settings, so teams avoid re-entering context each session.
Ramp engineers use OpenAI's Codex with GPT-4.5 to conduct code reviews, reducing feedback time from hours to minutes. The company has integrated the tool into its development workflow to automate substantive code analysis.
A developer published AgentOS 2.0, a collection of 135 structured Claude prompt "Skills" built over six months, each incorporating named sub-agents, domain-specific formulas, and runnable Python code rather than generic persona or instruction-based prompts.
Vercel's Chat SDK added a built-in AI SDK toolset accessible via a new `chat/ai` subpath, with a `createChatTools()` function that connects read and write actions to agents. Write tools require approval by default, and three presets — reader, messenger, and moderator — scope the available toolset.
Google announced Antigravity 2.0 at Google I/O 2026, expanding the platform from a coding environment into a system for developing and managing autonomous AI agent teams. The update includes a standalone desktop application, a CLI, and an SDK co-optimized for Gemini 3.5 Flash.
A developer released "claudenv," a shell-based account manager for Claude Code that uses the `CLAUDE_CONFIG_DIR` environment variable to switch between multiple Claude accounts. It supports global defaults, per-project `.claudenvrc` files, and auto-switching on directory change, requiring only ba...
A developer guide outlines use cases for Anthropic's Claude AI in web and e-commerce development, including React component generation, Next.js projects, Shopify store development, SEO optimization, and debugging across frontend and backend workflows.
A software engineer at Flower Shop Network used Claude to migrate a CI/CD pipeline from GitLab to AWS CodeBuild in 12 hours after GitLab's pricing structure made small top-ups impractical. Claude then authored a retrospective identifying five process mistakes made during the migration, including ...
Vercel's Chat SDK added support for `callbackUrl` props on buttons and modals, enabling Workflow runs to pause and resume upon user interaction. The feature works for buttons on most platforms with official adapters, and for modals on Slack and Teams.
A developer rebuilt a financial portfolio Q&A system using retrieval-augmented generation, finding that indexing real-time price data caused stale portfolio value errors and that vocabulary mismatch between test queries and real user language dropped context recall from 0.89 to 0.58.
Two students built a volleyball tournament tracking website using ChatGPT for planning, v0 by Vercel for UI generation, Supabase for the backend, and Claude Desktop to parse and restructure existing tournament spreadsheet data.
Mike Veerman built a browser-based tool that simulates LLM text output at speeds ranging from 5 to 800 tokens per second, letting users see what advertised token speeds look like in practice.
CodeGraph is an open source tool that pre-indexes codebases into a local SQLite semantic graph using tree-sitter, exposing 8 query tools to AI coding agents via MCP. Benchmarks across 7 projects show average reductions of 35% in token cost and 70% in tool calls; the project has 9,600+ GitHub stars.
Developer SaravananJaichandar released world-model-mcp v0.7.2, a Python package that gives AI coding agents persistent memory via SQLite, enforcing learned constraints at the edit boundary and flagging regressions against recorded bug fixes across Claude Code sessions.
xAI's Grok Build 0.1, a beta coding model designed for agentic coding tasks, is now available on Vercel AI Gateway. The model, which powers the Grok Build CLI app, is in early access and can be called via the AI SDK using the identifier `xai/grok-build-0.1`.
Developer Shrijith Venkatramana released git-lrc, an open-source AI code reviewer that runs on each Git commit and generates summaries to help teams maintain institutional knowledge of AI-generated code. The tool implements an "awareness loop" — converting commits into daily and weekly reports to...
A December 2025 analysis of 470 open-source GitHub pull requests found AI co-authored code had 1.7x more major issues and 2.74x higher security vulnerability rates than human-written code. BMAD is an open-source framework that applies traditional software development lifecycle phases to AI-assist...
Software architecture is shifting from deterministic input-output systems toward AI-native stacks that use language models as coordination layers between tools, workflows, and teams. The emerging approach centers on context engineering, retrieval pipelines, and trust mechanisms rather than model ...
Google announced Gemini Spark, a personal AI agent integrating with Gmail, Calendar, Drive, and other Google apps, running on Gemini 3.5 Flash and a new closed-source Go-based platform called Antigravity. Google also announced the open-source Gemini CLI will stop working with AI subscription plan...
GitHub disclosed that an attacker compromised an employee device via a malicious third-party VS Code extension, exfiltrating data from GitHub-internal repositories. The attacker claims approximately 3,800 repositories were accessed; GitHub found no evidence of impact to customer-owned data and ro...
Google announced approximately 100 products and updates at its I/O 2026 developer conference, including Gemini Omni, Google Antigravity, and Universal Cart.
NanoCo, a Tel Aviv startup, launched a managed enterprise AI agent service that deploys individual Docker-sandboxed agents for each employee, alongside a $12 million seed round led by Valley Capital Partners with Docker and Vercel participating. The company's open source NanoClaw framework has ac...
The 2026 WebAIM Million report found 95.9% of the top million homepages have detectable WCAG accessibility failures, with the average page containing 297 issues. AudioEye's chief accessibility officer attributes the trend partly to LLMs trained on inaccessible web content, which reproduce structu...
Vercel released a WordPress plugin that connects sites to its AI Gateway, providing access to models from 40+ providers including Anthropic, OpenAI, and Google via a single API key. The plugin requires WordPress 7.0, also released the same day, and supports text, image, and video generation.
Between January and April, six AI companies—Anthropic, Perplexity, Microsoft, OpenAI, Google, and Amazon—each launched agent products targeting knowledge workers, all sharing a common design: reading local files, controlling browsers, retaining context, and delivering completed outputs rather tha...
A Dev.to essay describes a divide in software development between developers who rely heavily on AI coding tools to accelerate output and those who reject them entirely, arguing both approaches carry legitimate risks depending on the operator's existing expertise.
At I/O 2026, Google announced AI agents capable of gathering information, planning events, and summarizing email and calendar data, with the ability to run continuously in the background. The announcements came as the company positioned itself as a leading contender in the AI agent space alongsid...