Agent infrastructure As of 2026-10-27
From terminal-based CLI agents to execution isolation, concurrency locks, and multi-agent orchestration. Every row below is an existing repository, with 30-day commit counts for the three core projects.
Rust TypeScript Go Swift Shell Python
The three repositories with the most development activity on the coding-agent track. Commit counts for the past 30 days are aligned on the right for a quick comparison.
A DeepSeek-native coding agent. Its Rust runtime, terminal UI, and release pipeline are tuned alongside the model's capabilities.
My main focus right now. Twenty blog posts cover its context budgets, empty-turn recovery, subagent delegation, permission sandbox, terminal UI redesign, and Terminal-Bench evaluations.
A CLI coding agent with 20+ built-in tools, MCP support, and routing across multiple models.
The user-facing product in the Blade family, and the first entry point for most users.
BMAD-style multi-agent orchestration, with distinct roles for planning, execution, QA, and delivery across a complete pipeline.
My most-starred repository. It turns multi-agent collaboration from a concept into a reproducible pipeline.
The foundation that lets those agents run safely: SDKs, execution isolation, concurrency locks, and observability.
A TypeScript SDK for building agents, with multi-turn conversation loops, tools, memory, and model-provider adapters.
A secure execution runtime (BAR) with git worktree isolation, a step ledger, and a policy engine.
Symbol-level locking and conflict prevention for parallel agents. Language-independent and ready to integrate.
A lightweight, agent-neutral macOS computer-use MCP server for Codex, Claude Code, and Orca.
The STOP protocol makes agent-skill execution observable, assertable, and debuggable.
Methods rather than runtimes: connecting requirements, design, execution, and retrospectives into a complete cycle.
A spec-driven development workflow connecting requirements, design, planning, and execution through living documentation.
A structured learning system for AI: knowledge graphs, detailed notes, practical retrospectives, and weekly reviews.
"Design Patterns for Vibe Coders": 17 chapters and 2 appendices. Follow a quick-note app as it grows from 44 lines to six files, learning to recognize patterns in AI-generated code, decide whether an abstraction is worthwhile, and remove it when needed.
A separate direction from Blade: make rendering fast, then let agents understand and edit the canvas.
A high-performance 2D rendering engine and canvas SDK. WebGPU instancing renders 100,000 rectangles in one draw call, with a snapshot + ops protocol that lets agents read and edit the canvas. Version 0.1.0 shipped in August (@sky-canvas/renderer and @sky-canvas/sdk).
Long-term projects with a stronger public-interest focus, presented separately from the technical work above.
Public interest · Real-world questions jiaoyuan.app A public-interest Q&A site for Jiaoyuan quotations, connecting classic texts, RAG retrieval, and real-world questions to help people find answers in the original sources.Small tools that are already live and ready to download or use, each with its own website.
A deep-cleaning tool for macOS: rule-based junk scans, storage analysis, large files, similar photos, and app uninstallation. Only confirmed junk is deleted; files in use are always skipped.
A desktop browser for ACP agents. Connect Claude Code, Codex, and other agents locally or over SSH, let them drive the browser through MCP, approve actions first, and take over at any time.
A quiet focus timer for macOS. Aquatic plants grow after each focus session, with occasional fish visitors. Includes a visitor collection, 20 ambient sounds, and a desktop widget.
Clipboard history for Mac: remembers copied text, links, code, and images. Press Cmd+Shift+V to select an entry and paste it back where you were working. History stays on your device.
A tool to preview, format, and publish Markdown as X Articles, so you can check a long-form post before publishing.
An AI text-and-image generator for Xiaohongshu, WeChat, and Douyin creators. Enter one sentence to generate topics, 3:4 information cards, titles, hashtags, and publishing copy.