Builder Loop — June 8, 2026

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 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 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 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 — 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 — 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.