// edition · 2026-06-03
30 stories on AI dev tools, agents, and the coding stack, curated from the day's RSS haul by Agentic Dev's pipeline.
Anthropic shipped Dynamic Workflows as a research preview in Claude Code alongside Claude Opus 4.8, enabling the tool to generate JavaScript orchestration scripts that spawn up to 1,000 parallel subagents and run adversarial "challenger" agents to verify results. A cited case study involved porti...
GitHub launched a desktop app called GitHub Copilot in technical preview at Microsoft Build, offering a unified interface to manage multiple AI coding agents running in parallel across repositories. The app is available to existing Copilot Pro, Pro+, Business, and Enterprise subscribers, with eac...
In Claude Code, the `description` field of a subagent configuration file serves as a routing signal for a dispatcher that decides whether to invoke the agent, not as human-readable documentation. Descriptions written as identity statements fail to trigger agents; descriptions listing literal trig...
Microsoft released a beta of Intelligent Terminal for Windows 11, which integrates AI coding agents directly into the shell environment. The feature supports Agent Communication Protocol-compatible agents including GitHub Copilot, Claude Code, and Codex, and can detect errors and surface agent su...
GitHub replaced Copilot's premium request model with token-based billing on June 1, 2026. Plan prices remain unchanged: Pro subscribers ($10/month) receive $15 in monthly AI credits, while Pro+ subscribers ($39/month) receive $70, with credits consumed based on token usage.
A developer published a Python template that replicates Claude Code's dynamic workflow pattern, splitting complex coding tasks into sequential role packets—planner, implementer, adversarial reviewer, and verifier—with trace logging and verification gates. The tool, `workflow_orchestrator.py`, is ...
A developer built "Gatekeeper," an open-source TUI daemon that intercepts Claude Code tool calls across multiple sessions and routes them to a single terminal dashboard for approval. The tool blocks certain commands—including rm, sudo, git push --force, and terraform destroy—from auto-approval re...
A guide published on Dev.to argues that AI coding agent failures typically stem from poorly structured repositories rather than inadequate prompts, recommending developers add explicit instruction files such as AGENTS.md covering test commands, package managers, and code boundaries. The author ci...
A developer proposes storing Claude Code context as individual single-fact Markdown files with YAML frontmatter, rather than one large CLAUDE.md file, to allow selective recall and easier deletion of stale entries. The approach uses an index file loaded each session while individual memory files ...
AI agents running in terminals lose all session context when closed because most architectures store state only in the running process. A common fix involves writing decisions and plans to files like daily logs and a bootstrap config, then using a cron job to reload that context on restart.
Microsoft released Adaptive Spec-driven Scoring for Evaluation and Regression Testing, an open source framework that lets developers create AI behavior evaluations using text descriptions.
Microsoft introduced a specification enabling developer, compliance, and security teams to define behavioral policies for AI agents using portable policy files. The system is designed to give organizations more granular control over how agents operate.
OpenAI announced new Codex plugins, site integrations, and annotation features aimed at analysts, marketers, designers, and investors, expanding the tool's use across professional workflows.
AI agents are being deployed in enterprise software to automate complex reasoning tasks, shifting development focus toward orchestrating autonomous systems. Key considerations for adoption include governance, scalability, and security integration.
A developer released Scribefy, an MCP server distributed as an npx package that gives Claude, Cursor, and Windsurf direct access to YouTube transcripts, metadata, search, and related videos via four tools, eliminating manual copy-paste workflows.
Microsoft announced two new LLMs: MAI-Thinking-1, a 1-trillion-parameter MoE reasoning model with 35B active parameters, and MAI-Code-1-Flash, a 137B-parameter model with 5B active parameters built for GitHub Copilot in VS Code. Both were trained from scratch on web crawl data including a proprie...
Microsoft announced Rayfin, an open-source SDK and CLI, at Build 2026 that allows developers and coding agents to define application backends in code and deploy them directly to Microsoft Fabric. Replit is the exclusive launch partner, with its agent using Rayfin to generate backends including da...
A benchmark test of three Claude Opus model versions (4.6, 4.7, 4.8) via the Crazyrouter API found Opus 4.7 achieved the highest pass rate (5/6 tasks) at 7.46s average latency, while Opus 4.8 posted the lowest latency at 4.59s with a 4/6 pass rate.
Microsoft launched Project Polaris at Build 2026, an in-house mixture-of-experts AI model that will replace GPT-4 Turbo as GitHub Copilot's default engine in August 2026, running on Azure's Maia accelerators. NVIDIA released Cosmos 3, a 16B–64B multimodal model trained on 20 trillion tokens handl...
Microsoft announced MAI-Thinking-1 at Build 2026, its first in-house advanced reasoning model, described as medium-sized and trained from scratch without distillation from third-party models. The company says it matches leading models on software engineering benchmarks.
Pragmatic Engineer author Gergely Orosz presented a keynote at Craft Conference in Budapest arguing that software engineers should slow down when working with AI coding agents to achieve better results. The talk coincides with near-universal adoption of AI coding tools among software engineers.
A developer released "wasmicro," a WebAssembly transformer inference runtime for browsers with a ~94 KB bundle size, built as an alternative to larger tools like Transformers.js and ONNX Runtime Web that can require hundreds of megabytes for simple text classification. It supports BERT encoder in...
NVIDIA and Microsoft announced RTX Spark with the N1X chip at GTC 2026, featuring 128GB unified memory aimed at running AI agents locally. Mininglamp released Mano-P, an open-source vision-driven GUI agent for Mac that runs without cloud API calls and achieves roughly 80 tokens/s on Apple M5 Pro ...
A developer published an API-based marketplace for AI agents where creators can register bots with a price-per-call fee, retaining 85% of revenue, plus a 5% perpetual referral cut. The platform claims to list over 1,000 bots across categories including trading and content analysis.
Microsoft opened its Build 2026 conference with announcements including the Majorana 2 quantum chip, a Surface mini PC aimed at AI developers, the Scout AI assistant built on OpenClaw, and Project Solara, an Android-based OS for AI agent devices. The company also released MAI Thinking 1, its firs...
Simon Willison released datasette-agent-micropython 0.1a0, an alpha plugin that runs Python code in a MicroPython/WebAssembly sandbox for Datasette Agent. The release is intended to allow safe code generation and execution within the Datasette Agent environment.
Anthropic expanded Project Glasswing, a cybersecurity initiative, to approximately 150 additional partner organizations that gain access to Claude Mythos Preview, an AI model the company says has identified thousands of high-severity vulnerabilities in major operating systems and browsers. The pr...
Microsoft launched Scout at its Build conference in San Francisco, a personal work agent built on OpenClaw and WorkIQ that integrates with Teams and Outlook to autonomously handle tasks such as scheduling conflicts and meeting preparation. Scout is currently available to Frontier customers, with ...
OpenAI announced two new Codex features — Sites and Annotations — targeting knowledge workers, who now make up 20% of the app's 5 million weekly active users. Sites lets users create shareable interactive dashboards and web apps, while Annotations extends document-editing capabilities to non-code...
Simon Willison released micropython-wasm 0.1a1, an alpha version of a WebAssembly-based MicroPython sandboxing library. The release includes fixes for limitations discovered while building the datasette-agent-micropython project.