# Builder's Daily — Builder Loop > Rolling 14-day signal for beat `builder-loop`. Ephemeral context — not evergreen corpus. > Author: Amit Kumar Agrawal | https://artificialcuriositylabs.ai > Generated: 2026-06-10 > Human index: https://artificialcuriositylabs.ai/daily/builder-loop/ > RSS: https://artificialcuriositylabs.ai/daily/builder-loop/rss.xml --- # Builder Loop — June 10, 2026 **URL:** https://artificialcuriositylabs.ai/daily/builder-loop/2026-06-10/ **Beat:** builder-loop **Date:** 2026-06-10 **Topics:** github-copilot, anthropic, model-release, ide-integration, github, security **Summary:** Claude Fable 5 reaches general availability across GitHub Copilot surfaces; GitHub extends automatic PR security validation to third-party coding agents… ## The read Coding agents are bulldozers, not replacements — humans still frame the problem. OSS shifts default stack choices faster than any vendor roadmap. When everyone can generate and fork, differentiation is taste, review, and what you ship before it becomes the default. ## What moved - **Claude Fable 5 reaches general availability across GitHub Copilot surfaces** — [GitHub Changelog](https://github.blog/changelog/2026-06-09-claude-fable-5-is-generally-available-for-github-copilot) Anthropic's Claude Fable 5 — described as the first model in Anthropic's 'Mythos' class, built for long-horizon autonomous coding and knowledge-work — is now GA in GitHub Copilot across VS Code, Visual Studio, Copilot CLI, the Copilot cloud agent, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse, for Pro+, Max, Business, and Enterprise plans. Unlike other Claude models in Copilot, Fable 5 requires 30-day data retention so Anthropic's safety classifiers can run, and Business/Enterprise admins must explicitly enable its policy. **Builder angle:** A new long-horizon coding model lands in every Copilot surface at once, but Business/Enterprise teams must opt into a 30-day data-retention policy before they can use it. - **GitHub extends automatic PR security validation to third-party coding agents like Claude and Codex** — [GitHub Changelog](https://github.blog/changelog/2026-06-09-security-validation-for-third-party-coding-agents) GitHub now runs its automatic security checks — CodeQL vulnerability analysis, dependency scanning against the GitHub Advisory Database, and secret scanning — on pull requests from any coding agent, not just Copilot's cloud agent. This covers third-party agents including Claude and OpenAI Codex. Agents attempt to fix flagged issues before finalizing the PR. The checks run by default, follow existing repo Copilot settings, and require no GitHub Advanced Security license. **Builder angle:** PRs opened by Claude, Codex, or other third-party agents now get the same CodeQL/secret-scan/dependency gate as Copilot's cloud agent, automatically and with no extra license. - **VibeDrift launches MCP server that feeds codebase-convention signals into Claude Code and Cursor mid-generation** — [VibeDrift Blog](https://www.vibedrift.ai/blog/does-a-drift-checker-change-agent-output) VibeDrift shipped a paid-tier MCP server that runs inside coding-agent sessions (Claude Code, Cursor) and surfaces codebase-convention signals during code generation, rather than scanning for drift after the fact. The team's own measurement found that when a project convention conflicts with a model's default and isn't already visible in context, the MCP signal measurably reduced introduced drift (95% CI [0.57, 1.11]) — with no effect when conventions matched defaults or were already in-file. **Builder angle:** One MCP config block can cut convention drift in agent-written code, specifically in the cases where a model's defaults disagree with house style and that style isn't already visible in context. ## Also tracking - **Cohere ships North Mini Code, a 30B MoE (3B active) coding model under Apache 2.0** — [source](https://huggingface.co/blog/CohereLabs/introducing-north-mini-code) — Apache 2.0 license plus a small 3B active-parameter footprint and FP8 checkpoints make this a realistic self-host target for terminal coding agents, not just an API-only release. - **Vercel AI Gateway data: DeepSeek jumped from <1% to 17% of token volume in a month, while spend share stayed near 1%** — [source](https://vercel.com/blog/ai-gateway-production-index-june-2026) — Concrete production routing data shows an open-weight model now carrying ~1/6 of gateway token volume at near-zero cost share — a real comparison point for teams deciding which open model to route bulk/cheap traffic to. --- # Builder Loop — June 9, 2026 **URL:** https://artificialcuriositylabs.ai/daily/builder-loop/2026-06-09/ **Beat:** builder-loop **Date:** 2026-06-09 **Topics:** xcode, apple, acp, mcp, wwdc, ios27 **Summary:** Apple ships Xcode 27 with native ACP and MCP support, integrating Claude, Gemini, and OpenAI coding agents into the IDE; Anthropic releases Swift packag… ## The read Coding agents are bulldozers, not replacements — humans still frame the problem. OSS shifts default stack choices faster than any vendor roadmap. When everyone can generate and fork, differentiation is taste, review, and what you ship before it becomes the default. ## What moved - **Apple ships Xcode 27 with native ACP and MCP support, integrating Claude, Gemini, and OpenAI coding agents into the IDE** — [Apple Newsroom](https://www.apple.com/newsroom/2026/06/apple-aids-app-development-with-new-intelligence-frameworks-and-advanced-tools/) Xcode 27, announced at WWDC 2026 on June 8, adds native Agent Client Protocol (ACP) and Model Context Protocol (MCP) support. Coding agents from Anthropic, Google, and OpenAI run directly inside the IDE with interactive planning, multiturn conversations, side-by-side code previews, and autonomous test/simulator validation via the new Device Hub. External tools like GitHub and Figma connect via MCP. The Foundation Models framework gains a single Swift API supporting server models and image input; Small Business Program developers get Apple on-device model access at no cloud API cost. **Builder angle:** Xcode 27 is now the first Apple IDE where any ACP-compatible coding agent runs natively and validates its own code against your simulator — no third-party plugin or manual handoff required. - **Anthropic releases Swift package for Claude + Apple Foundation Models, enabling on-device/cloud AI handoff in SwiftUI apps** — [Anthropic Blog](https://claude.com/blog/claude-for-foundation-models) Anthropic released a native Swift package that integrates Claude with Apple's Foundation Models framework for iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. The package accepts typed value outputs from Apple's on-device model and routes them to Claude for multi-step reasoning, code generation, or web search, returning streaming responses, tool calls, and structured data back into SwiftUI views. Developers work entirely in Swift without handling raw prompt text. **Builder angle:** Apple platform developers can chain on-device Foundation Model outputs directly into Claude with a single Swift import — typed input, streaming response, and tool calls included, without raw prompt management. - **JetBrains Junie CLI adds --acp flag, making it a protocol-native agent for any ACP-compatible editor** — [JetBrains Junie Docs](https://junie.jetbrains.com/docs/junie-cli-acp.html) JetBrains published documentation dated June 8, 2026 for Junie CLI's ACP mode. When launched with — acp true, the CLI shifts from interactive terminal mode to serving requests initiated by external ACP-compatible editors and IDEs over JSON-RPC via stdio or HTTP/WebSocket. In ACP mode Junie exposes diff generation, Markdown-formatted responses, and IDE state inspection to any compliant host editor, decoupling the JetBrains agent from a single IDE host. **Builder angle:** Any editor that speaks ACP can now delegate to Junie without a JetBrains-specific plugin — the same protocol Xcode 27 and Devin Desktop adopted, making Junie a drop-in agent for polyglot editor environments. ## Also tracking - **verl v0.8.0 ships as ByteDance's production-grade open RL training library with vLLM/SGLang integration** — [source](https://github.com/volcengine/verl) — Builders running RLHF or RL-from-feedback pipelines can drop verl into any vLLM- or SGLang-backed cluster and get a production-tested PPO/GRPO loop that eliminates the training-inference handoff bottleneck. - **OpenEnv launches as community-governed open standard for agentic RL environments, backed by Meta-PyTorch, Nvidia, Unsloth, and Hugging Face** — [source](https://huggingface.co/blog/openenv-agentic-rl) — Builders training agents via RL can write one OpenEnv-compliant environment and plug it into any trainer (verl, prime-rl, VeRL-Omni) without rewriting environment adapters per framework. --- # Builder Loop — June 8, 2026 **URL:** https://artificialcuriositylabs.ai/daily/builder-loop/2026-06-08/ **Beat:** builder-loop **Date:** 2026-06-08 **Topics:** ACP, IDE, agent-orchestration, Cognition, Windsurf, code-review **Summary:** Cognition rebrands Windsurf as Devin Desktop, ships native Agent Client Protocol support; Snap details CodePal, an AI code reviewer that runs a multi-pa… ## The read Coding agents are bulldozers, not replacements — humans still frame the problem. OSS shifts default stack choices faster than any vendor roadmap. When everyone can generate and fork, differentiation is taste, review, and what you ship before it becomes the default. ## What moved - **Cognition rebrands Windsurf as Devin Desktop, ships native Agent Client Protocol support** — [Cognition Blog](https://cognition.ai/blog/introducing-devin-desktop) Cognition relaunched Windsurf as Devin Desktop, making the Agent Command Center (a Kanban board of local and cloud agents, sorted by status) the default surface, and shipping native support for the open Agent Client Protocol so any ACP-compatible agent — Codex, Claude Agent, OpenCode, or in-house builds — can run inside the editor alongside Devin. The change rolled out as an over-the-air update on June 2, 2026; existing Windsurf plans, pricing, and extensions carry over. **Builder angle:** ACP support means teams can standardize on one editor while running whichever agent fits a given task, breaking the lock-in between IDE and agent vendor. - **Snap details CodePal, an AI code reviewer that runs a multi-pass verification loop on every PR** — [Snap Engineering Blog](https://eng.snap.com/codepal) Snap published the architecture behind CodePal, its mandatory pre-human PR reviewer: two parallel bootstrap passes with different sampling parameters, a speculative third pass on disagreement, additional passes only when new findings emerge, and a verifier that strips hallucinated or contradictory findings before posting. It builds context without cloning repos — using GitHub Enterprise APIs and tree-sitter symbol indexing — and does cross-repo semantic search to catch downstream breakages. Snap reports 200,000+ reviews over four months at ~$0.40/review, finishing in ~10 minutes versus ~5 hours for first human review, recall climbing from 30% to 80%, and 90% PR adoption within a quarter (up from a 9% pilot). **Builder angle:** The multi-pass-plus-verifier pattern and concrete cost/recall numbers ($0.40/review, 30%→80% recall) give teams a reproducible blueprint for replacing or gating human first-pass review with an agent. - **GitHub Copilot in Visual Studio adds a Plan agent that drafts implementation plans before code is written** — [GitHub Changelog](https://github.blog/changelog/2026-06-04-github-copilot-in-visual-studio-may-update) The May update to Copilot in Visual Studio ships a new Plan agent that analyzes the codebase with read-only tools, asks clarifying questions, and produces a markdown implementation plan that can then be handed to Agent mode for execution. The release also adds a Skills panel for managing discovered agent skills, a multi-file change-review summary view (accept/reject by file, all-files, or chunk), a context-window usage ring with conversation summarization, and the ability to attach Git History/Blame commits directly as chat context. **Builder angle:** Splitting 'plan' from 'execute' into separate agent modes gives developers a checkpoint to review and edit scope before an agent starts editing files — directly changes the review-before-you-build loop. ## Also tracking - **NVIDIA releases Nemotron 3 Ultra as open-weight, open-data, and open-recipe under OpenMDW-1.1 with reproducible agentic benchmarks** — [source](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents) — Builders can fork the full training recipe (not just run inference) and reproduce NVIDIA's published agentic-coding and long-context benchmark numbers from the same Hugging Face checkpoints. - **Google ships Gemma 4 12B, an encoder-free multimodal model that runs locally on a 16GB-VRAM laptop GPU** — [source](https://developers.googleblog.com/gemma-4-12b-the-developer-guide/) — A 12B multimodal model that fits a consumer GPU and lands directly in Ollama/LM Studio pipelines lowers the bar for local agentic prototyping with combined audio, vision, and text inputs. --- # Builder Loop — June 7, 2026 **URL:** https://artificialcuriositylabs.ai/daily/builder-loop/2026-06-07/ **Beat:** builder-loop **Date:** 2026-06-07 **Topics:** github-copilot, ci-cd, cloud-agent, agentic-workflow, vs-code, model **Summary:** Fix with Copilot for failing Actions now in Pro, Pro+, and Max; MAI-Code-1-Flash is now available for GitHub Copilot in VS Code; JetBrains Mellum2 goes … ## The read Coding agents are bulldozers, not replacements — humans still frame the problem. OSS shifts default stack choices faster than any vendor roadmap. When everyone can generate and fork, differentiation is taste, review, and what you ship before it becomes the default. ## What moved - **Fix with Copilot for failing Actions now in Pro, Pro+, and Max** — [GitHub Changelog](https://github.blog/changelog/2026-06-04-fix-with-copilot-for-failing-actions-now-in-pro-pro-and-max) When a GitHub Actions job fails, a 'Fix with Copilot' button on the workflow log triggers a cloud agent that investigates the failure, implements a fix, pushes it to the branch, and notifies the developer for review—no manual triage required. Available on Copilot Pro, Pro+, and Max. **Builder angle:** Builders on GitHub Actions can now hand CI failure triage and patching to a cloud agent, removing the context-switch to debug broken runs before resuming feature work. - **MAI-Code-1-Flash is now available for GitHub Copilot in VS Code** — [GitHub Changelog](https://github.blog/changelog/2026-06-02-mai-code-1-flash-is-now-available-for-github-copilot/) Microsoft's in-house small-tier coding model, MAI-Code-1-Flash, is rolling out in GitHub Copilot via the VS Code model picker. Outperforms comparable small models in early testing and is available across all Copilot tiers (Free through Max), starting with a limited rollout expanding over weeks. **Builder angle:** VS Code Copilot users now have a Microsoft-optimized, low-latency coding model selectable from the model picker—a cheaper inference option for routine completions without leaving the IDE. - **JetBrains Mellum2 goes open source: 12B MoE model for coding agent routing and sub-agents** — [JetBrains Blog](https://blog.jetbrains.com/ai/2026/06/mellum2-goes-open-source-a-fast-model-for-ai-workflows/) JetBrains released Mellum2, a 12B-parameter Mixture-of-Experts model (2.5B active per token) under Apache 2.0 on Hugging Face. Designed for routing, low-latency RAG, and sub-agent orchestration in coding pipelines—inference time reportedly less than half of similar-sized dense models. Supports private, local deployment. **Builder angle:** Teams building coding agent pipelines or IDE integrations can self-host a JetBrains-tuned model optimized for context-gathering, planning, and validation steps at a fraction of frontier model inference cost. ## Also tracking - **OpenCV 5.0 ships rewritten DNN engine with built-in LLM/VLM inference and 80%+ ONNX coverage** — [source](https://github.com/opencv/opencv/releases/tag/5.0.0) — CV pipelines can now run VLM inference (image→text) in-process via OpenCV without a separate LLM runtime, enabling tighter perception–language integration in agent and robotics deployments. - **Hugging Face redesigns `hf` CLI as agent-first tool: dual-mode output, next-command hints, 94% task success in Claude Code** — [source](https://huggingface.co/blog/hf-cli-for-agents) — Agent harnesses can call Hub operations (model download, Space deploy, dataset push) via structured CLI output with measurably higher success rates and lower token cost than SDK or curl wrappers. --- # Builder Loop — June 6, 2026 **URL:** https://artificialcuriositylabs.ai/daily/builder-loop/2026-06-06/ **Beat:** builder-loop **Date:** 2026-06-06 **Topics:** acp, cli, windows-terminal, copilot-cli, cursor-sdk, subagents **Summary:** Intelligent Terminal 0.1 ships ACP-native agent pane in Windows shell; Cursor SDK adds custom tools, nested subagents, and headless auto-review; GitHub … ## The read Coding agents are bulldozers, not replacements — humans still frame the problem. OSS shifts default stack choices faster than any vendor roadmap. When everyone can generate and fork, differentiation is taste, review, and what you ship before it becomes the default. ## What moved - **Intelligent Terminal 0.1 ships ACP-native agent pane in Windows shell** — [Windows Command Line Blog](https://devblogs.microsoft.com/commandline/announcing-intelligent-terminal-version-0-1/) Microsoft released Intelligent Terminal 0.1, an experimental Windows Terminal fork with a docked agent pane connected via Agent Client Protocol (ACP). GitHub Copilot CLI is the default agent, but any ACP-compatible CLI is configurable. Failed commands trigger auto-detected errors in the status bar; clicking loads shell output into the agent pane for explanation and fixes. Background agent tasks run in new tabs, and Command Palette prompt mode (?query) injects active-pane context without blocking the shell. **Builder angle:** Terminal-first developers can delegate multi-step shell fixes to any ACP agent without leaving the command line or manually copying error output. - **Cursor SDK adds custom tools, nested subagents, and headless auto-review** — [Cursor Changelog](https://cursor.com/changelog) Cursor's June 4, 2026 SDK release lets local agents register custom tools via function definitions exposed through a built-in custom-user-tools MCP server, visible to all nested subagents. Subagents can spawn subagents to arbitrary depth. Headless runs can set local.autoReview to route tool calls through a classifier steered by permissions.json allow/block instructions. Persistence options expand beyond SQLite to JSONL and custom LocalAgentStore implementations; each send() carries a platform requestId for CI correlation. **Builder angle:** CI and internal scripts can embed coding agents with first-class custom tools and graded auto-approval instead of standing up separate MCP servers or interactive review loops. - **GitHub Agent tasks REST API exposes programmatic Copilot cloud agent runs** — [GitHub Changelog](https://github.blog/changelog/2026-06-04-agent-tasks-rest-api-now-available-for-copilot-pro-pro-and-max/) Copilot Pro, Pro+, and Max users can start and track Copilot cloud agent tasks via a public-preview REST API authenticated with PATs or OAuth tokens. Cloud agents run in an isolated development environment, make and validate code changes, and open pull requests. Documented use cases include fan-out refactors across repositories, one-click repo scaffolding from internal portals, and scheduled release preparation with release notes. **Builder angle:** Background coding agents can be triggered from scripts, portals, or schedulers instead of only from IDE or Copilot app sessions. ## Also tracking - **NVIDIA open-sources physical AI agent skills across Omniverse, Cosmos, Alpamayo, and Metropolis** — [source](https://nvidianews.nvidia.com/news/nvidia-releases-major-collection-of-open-source-agent-tools-and-skills-for-physical-ai) — Agent builders working on embodied or simulation-heavy workflows can pull verified NVIDIA skills into existing harnesses instead of wiring CUDA-X libraries by hand. - **vLLM Semantic Router adds Session-Aware Agentic Routing with prefix-cache switch pricing** — [source](https://vllm.ai/blog/2026-06-02-session-aware-agentic-routing) — Self-hosted agent stacks using vLLM auto routing can keep multi-turn tool sessions stable without silently breaking provider continuation state or wasting prefix-cache locality.