// edition · 2026-06-13
32 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
Vercel released AI SDK 7 with a `HarnessAgent` API that provides a unified interface for running agent harnesses including Claude Code, Codex, and Pi. The experimental feature, available on the canary release, normalizes access to harness capabilities such as sandboxes, sessions, and permission f...
Claude Code skills are markdown instruction files stored in ~/.claude/skills/ that define autonomous agent behavior, including stop-loss conditions such as token budgets and deadlines. A developer testing autonomous agents on GitHub translation bounties found that explicit guardrails and decision...
A developer described using Cursor for local, in-editor coding tasks roughly 90% of the time, and Anthropic's Claude Code for codebase-wide, multi-file tasks the remaining 10%. The distinction drawn is that Cursor functions as an inline coding assistant, while Claude Code operates as a terminal-n...
Anthropic will move Claude's headless mode (`claude -p`) and Agent SDK to a separate metered credit system on June 15, separating it from flat-rate subscription access. GitHub Copilot made a similar shift on June 1, and OpenAI Codex uses comparable split pricing between seat costs and API usage.
A developer published a guide on configuring a locally-run AI coding agent on macOS, covering the tools and steps required to run the setup without relying on cloud-based services.
A developer used a Claude model codenamed "Fable 5" via Claude Code's /goal command to build a browser-based retro racing game in three.js, HTML5/CSS/JS, and open-source GLB car models. The model ran autonomously for over 30 minutes, built its own Python testing framework, and took screenshots to...
A developer built a Claude Code plugin called Chrome Extension Builder to validate Manifest V3 Chrome extensions, using five slash commands and three agents with distinct capability limits to catch MV3-specific errors like invalid CSP directives and over-broad permissions before browser or store ...
GitHub updated Copilot CLI (v1.0.42+) with smarter subagent delegation, making the tool more selective about when to hand tasks to helper agents. An A/B test showed a 23% reduction in tool failures per session and up to 5% improvement in user wait time at P95, with no quality regression.
A staff engineer describes replacing traditional ETL pipelines with an "agentic data serving" architecture that uses DuckDB, S3 storage, and the Model Context Protocol (MCP) to allow AI agents to query data directly without hand-coded transformation steps.
A developer building an AI governance system called CORE found that across 35,669 logged LLM calls, the cost_estimate column was never populated, leaving all cost data as NULL. The system recorded token counts, model usage, and timestamps but could not attribute costs to specific AI roles or mode...
An AI agent tasked with scanning DN42, a decentralized hobbyist network, generated costs large enough to bankrupt its operator by making an uncontrolled volume of requests or API calls during the scan.
Vercel launched Drop, a browser-based deployment tool at vercel.com/drop that lets users deploy files or folders by dragging them into the browser without Git or CLI setup. It supports framework projects such as Next.js and static site exports from tools including Bolt.new, Claude Design, and Goo...
A developer built a tool to verify that refactored functions produce identical outputs by running both versions against the same inputs, but encountered non-reproducible results when testing floating-point math — the same code returned different verdicts on macOS versus Linux CI due to IEEE-754 a...
Workflow SDK released a beta integration with Nitro v3, allowing workflow steps to run within the same bundled runtime as the rest of the app. The integration includes access to Nitro server APIs inside step functions and a workflow monitoring UI served at `/_workflow`.
A developer tutorial walks through training a character-level language model on Shakespeare's complete works (~1MB of text) using Andrej Karpathy's nanoGPT repository, targeting consumer hardware such as Apple Silicon Macs. The process takes approximately 15 minutes and produces text that mimics ...
Anthropic released Claude Fable 5 on June 9, 2026, the public version of its previously internal Mythos-class model, which had been withheld due to its ability to find software vulnerabilities. An evaluation of ~1,000 agent tasks found Opus 4.8 still offers better value for most production coding...
Claude Desktop supports Model Context Protocol (MCP) servers, which connect the app to external tools including file systems, GitHub repos, databases, and web search via a JSON config file. The setup requires at least a Claude Pro subscription ($20/month) and is available natively on macOS and Wi...
Researchers have found that 43% of public MCP servers contain at least one vulnerability, and 5.5% have poisoned tool descriptions that can silently redirect AI agent behavior in editors like VS Code and Cursor. OX Security disclosed in May that the official MCP SDK's local transport could be exp...
A developer released "Tribunal," an open-source Claude skill that reviews code diffs using multiple adversarial sub-agents assigned distinct roles — a critic, an integration checker, a judge, and a verdict compiler — rather than a single model. The tool is MIT licensed, requires no external depen...
Moonshot AI's Kimi K2.7 Code, a coding model supporting text and vision input that runs in thinking mode, is now available on Vercel's AI Gateway via the identifier `moonshotai/kimi-k2.7-code` in the AI SDK.
Moonshot AI released Kimi K2.7-Code, an open-source coding model available on HuggingFace, designed with improved token efficiency compared to existing coding models.
Anthropic suspended access to two Claude models, Claude Mythos 5 and Claude Fable 5, according to an incident report on the company's status page.
Vercel suspended access to Claude Fable 5 on its AI Gateway platform following a US Government directive, with no timeline given for restoration. Other Anthropic models remain available through the service.
Stack Overflow's monthly question volume fell from a peak of ~289,000 in early 2014 to the low thousands by end of 2025, following the launch of ChatGPT. The company responded by launching a product aimed at serving AI coding agents, entering a space where Mozilla and Andrew Ng had already releas...
Anthropic reportedly disabled access to Claude Mythos 5 and Fable 5 for all customers on June 12, 2026, following a US government export-control directive tied to national security concerns over foreign-national access.
Simon Willison wrote about Claude Fable, describing the AI agent as persistently taking initiative and acting without explicit prompting. Fable appears to be an agentic product built on Anthropic's Claude model.
The US government issued an export control directive requiring Anthropic to suspend all customer access to its Fable 5 and Mythos 5 AI models, citing national security concerns about a reported jailbreak technique. Anthropic received the directive at 5:21pm ET on June 12, 2026, and access was cut...
Simon Willison updated his browser-based OpenAI WebRTC audio tool to support the GPT-Realtime-2 model and added an optional document context field, allowing users to paste text and discuss it via voice in the browser.
The US government issued an export-control directive on June 12 requiring Anthropic to cut off access to its Claude Fable 5 and Mythos 5 models for foreign nationals, prompting Anthropic to suspend access globally. Anthropic stated it disagreed with the directive and was working to restore access.
The US government issued an export control directive Friday ordering Anthropic to disable its Fable 5 and Mythos 5 AI models, citing national security concerns over an alleged jailbreak, just three days after launch. Anthropic complied by pulling the models for all customers, disputing the govern...
A system architect with ERP and network infrastructure experience outlines the differences between traditional system architecture and AI solution architecture, noting the latter shifts focus from hardware/software layers to data quality, model selection, and prompt engineering.
A developer claims Anthropic's "Claude Fable 5" model launched June 9 and was pulled three days later, on June 12, following a reported U.S. government export control directive. The author generated 2,889 lines of code for a Figma-style canvas editor in a 28-minute, $9.54 session before the model...