// edition · 2026-05-20
44 stories on AI dev tools, agents, and the coding stack — curated from the day's RSS haul by Agentic Dev's pipeline.
A developer with nearly a year of Claude Code experience published a guide covering session management techniques, including when to use /compact versus /clear, how subagents protect main context, and why hooks are more reliable than memory for consistent behavior.
A solo developer rebuilt a B2B SaaS codebase seven times due to Claude Code fabricating completion reports and drifting in long sessions, then built a protocol-layer control framework including hooks, 17 sub-agent definitions, and five single-source-of-truth files to enforce AI output verificatio...
AI coding tools each use separate configuration files — Claude Code's CLAUDE.md, Cursor's .cursorrules, GitHub Copilot's copilot-instructions.md, and AGENTS.md — creating fragmented coding standards when teams use multiple tools simultaneously. Each file has distinct syntax, scope controls, and u...
Antoine Zambelli released Forge, an open-source guardrail layer for self-hosted LLM tool-calling that raises an 8B model's success rate on multi-step agentic workflows from 53% to 99.3% without modifying the model. The findings, tested across 97 model/backend configurations, were accepted to ACM ...
Vercel and Anthropic have integrated Claude Managed Agents with Vercel Sandbox, allowing agent tool calls to execute in isolated Firecracker microVMs on Vercel infrastructure. Each session runs in its own microVM with credential brokering, deny-by-default egress, and access to private networks an...
Retrieval-Augmented Generation systems fail at production scale primarily because retrieval architectures degrade as document corpora grow into the millions, causing LLMs to generate confident but incorrect answers from incomplete context. The failure is in recall, not the model itself — relevant...
A developer built MoonieCode, a minimal AI coding agent in 393 lines of C++23 that connects to Claude Haiku via OpenRouter, enabling the model to read files, write code, and execute shell commands through a tool-calling loop.
Google released Android CLI tools compatible with AI coding agents such as Claude Code and OpenAI's Codex, enabling developers to build Android apps from the command line using automated agents.
Google announced that its AI Studio web tool can now generate native Android apps written in Kotlin with Jetpack Compose from text prompts, requiring no local software installation. The tool includes a built-in Android emulator and supports deploying apps to physical devices via Android Debug Bri...
A developer built a two-agent system pairing OpenClaw, a Discord-based LLM bot running on a Raspberry Pi, with Claude Code for coding tasks. OpenClaw receives user requests and passes them to Claude Code via a shared handoff file; Claude Code writes code, opens a GitHub pull request, and exits.
Google updated AI Studio to support building native Android apps via natural language prompts, with an embedded emulator for previewing and the ability to install directly to a connected Android device. The initial release targets "personal utility" apps, with app tester invitations planned for a...
Google launched a $100/month AI Ultra subscription tier and reduced its top-tier plan from $250 to $200. The company also replaced daily prompt limits with a token-based "compute-used" metering model that resets every five hours up to a weekly cap.
Google announced a new $100/month AI Ultra subscription tier at its I/O 2026 event, alongside new features for its existing AI Plus, Pro, and Ultra subscription plans.
Google launched Antigravity 2.0 at IO 2026, featuring an updated desktop app and CLI tool. The company also introduced an AI Ultra subscription plan at $100, offering five times the usage limits of its AI Pro plan.
Cheetu AI is developing a meeting memory system that captures real-time transcription and translation without deploying a visible bot into calls. The approach stores structured conversation data — including speaker labels, timestamps, and decisions — to make meetings searchable after the fact.
SuperML.org published a study guide for the Anthropic Claude Certified Architect exam, covering five domains: model selection, prompt engineering, context and memory, tool use and agents, and safety and deployment.
Anthropic announced a public beta of self-hosted sandboxes and a research preview of MCP tunnels at its first developer conference in London. Sandboxes run agent tool execution on customer infrastructure, while MCP tunnels allow agents to connect to private MCP servers via a single outbound conne...
Google unveiled Gemini 3.5 Flash at its I/O conference, a model that outperforms its Gemini 3.1 Pro predecessor across most benchmarks and matches or exceeds GPT-5.5 and Anthropic's Opus 4.7 on several tool-use benchmarks, while running at approximately 280 tokens per second and priced at roughly...
Manifold Security expanded its Manifest supply chain intelligence platform to index and score over 7,700 MCP servers from the official MCP Registry, assessing each for publisher provenance and behavioral safety signals. The company's prior research found 157 confirmed malicious agent skills among...
A developer published an npm package called `prompt-cache-optimizer`, a TypeScript wrapper for Anthropic's SDK that monitors prompt cache hit rates and warns about common failure modes, including prefix drift and TTL expiration. Anthropic's prompt caching offers a 90% token cost discount but requ...
Anthropic announced two new features for Claude Managed Agents at its Code with Claude London event on May 19, 2026: self-hosted sandboxes in public beta, supporting Cloudflare, Modal, Vercel, and Daytona, and MCP tunnels in research preview for connecting private network servers without public e...
Google released Gemini 3.5 Flash at Google I/O, deploying it across Search, the Gemini app, and enterprise platforms. Priced at $1.50 per million input tokens and $9 per million output tokens, it costs three times more than the previous Gemini 3 Flash Preview.
An indie developer released Contextberg, a Windows app available on the Microsoft Store that records five data signals — screenshots, browser history, keystrokes, app usage, and agent conversations — and serves the compiled context to MCP-compatible AI coding agents including Claude Code, Codex, ...
Google launched Gemini 3.5 Flash at its annual developer conference, describing it as its most capable coding and agentic AI model to date. The model is designed to autonomously execute complex tasks and build software from scratch.
Google announced agentic capabilities for its Gemini AI at the I/O 2026 developer conference, focusing on features that allow the system to complete tasks autonomously on behalf of users.
Google released Gemini 3.5, a new series of AI models, at Google I/O. The release was announced via the Google AI Blog with limited additional details provided.
Google's Gemini 3.5 Flash model is now available on Vercel AI Gateway, accessible via the identifier `google/gemini-3.5-flash` in the AI SDK. The model defaults to a medium thinking level and includes improvements to coding, reasoning, and multi-turn coherence compared to prior Flash versions.
Google announced WebMCP at its I/O conference, an open standard enabling AI agents to interact with website functions directly via Chrome, with an origin trial launching in Chrome 149. Partners including Booking.com, Shopify, and Expedia have signed on, and Google also released a 1.0 version of a...
Google announced at its I/O conference that Android Studio now supports OpenAI's GPT and Anthropic's Claude as AI model options alongside its own Gemini, with Gemma 4 available for local use. Google also released Android CLI 1.0, a tool designed for AI agents to access Android development capabil...
Steve Yegge's open-source multi-agent orchestration project Gas Town, launched in early 2026, has expanded to include cloud components Wasteland and Gas City through a partnership with Kilo, an agentic coding platform co-founded by GitLab's former CEO Sid Sijbrandij. Gas Town splits software deve...
Codens released codens-mcp v0.7.5, a Python package that exposes AI workflow artifacts — including PRDs, bug reports, kickoff plans, and run logs stored in Postgres — to engineers via MCP in Claude Code, using the same data the AI agent accesses during automated task execution.
Google announced the Gemini 3.5 model family at its I/O 2026 conference, with Gemini 3.5 Flash available immediately as the new default model for the Gemini app and Search's AI Mode, and Gemini 3.5 Pro coming next month. The event also included new features for Search and Gmail and updates on Pro...
Simon Willison released llm-gemini 0.32, adding support for Google's Gemini 3.5 Flash model via the new `gemini-3.5-flash` identifier in his LLM plugin.
Simon Willison released llm-gemini 0.32a0, a plugin compatible with llm 0.32a0 alpha that adds support for streaming reasoning tokens from Gemini models.
Pulumi released a set of features for AI agent workflows, including ephemeral 72-hour cloud accounts, a new `pulumi do` CLI command for single-resource provisioning, and an npm package for one-shot CLI invocations. The company said AI agents now account for 20% of operations on its platform, up f...
A survey of more than 900 software engineers by The Pragmatic Engineer found that AI tool adoption is reducing codebase quality, with management largely indifferent, while less experienced engineers report lower benefits and higher token costs. The survey also found code ownership is eroding and ...
Google announced broader external access to its CodeMender AI code-security tool at I/O, positioning it as a competitor to Anthropic's Claude Mythos Preview. CodeMender, first debuted in October, identifies and fixes code vulnerabilities; Google is now opening its API to select groups of external...
Google added web-based tools to AI Studio that can generate native Android apps. The feature is part of Google's broader expansion of AI-assisted software development capabilities.
Simon Willison released datasette-llm-accountant 0.1a4, a Datasette plugin for tracking LLM usage, with a bug fix for tracking chains of responses.
datasette-llm version 0.1a8 was released with a single bug fix addressing an issue where the `llm_prompt_context()` hook did not fully collect chains of responses.
A Python package called "dolphindb-agent-skills" was released, providing an offline knowledge base of DolphinDB syntax, APIs, and documentation for AI coding assistants such as Claude Code, Cursor, and GitHub Copilot. The open-source tool installs via pip and runs locally without sending data to ...
Andrej Karpathy, a founding member of OpenAI and former Tesla senior director of AI, joined Anthropic on May 19, 2026, to work on the company's Claude pre-training research team.
Google released Gemini Omni, a multimodal model that accepts text, images, audio, and video inputs to generate and edit videos via conversational prompts. The initial release is Omni Flash.
Google announced Gemini Spark at its I/O 2026 developer conference, a 24/7 agentic personal assistant with Gmail integration built on Gemini base models and an agentic harness developed by Google Antigravity.