Dev Digest — April 5, 2026
🔥 HOT RELEASES
Swarm 0.1 — Parallel Coding Agent Manager Linux desktop app (powered by Ghostty) for managing multiple coding agents in parallel. One workspace per feature, one agent per workspace — git worktrees, terminal sessions, no more tab chaos. Works with any coding agent. 🔗 X: https://x.com/penberg/status/2040693936113476075 💻 GitHub: https://github.com/penberg/swarm
OpenFang — Production-Grade Agent OS in Rust 137K lines of Rust, 14 crates, 1,767+ tests, zero clippy warnings. Cold starts under 200ms, 40MB idle memory. SQLite memory backend with vector embeddings. This is what serious agent infrastructure looks like. 🔗 X: https://x.com/techNmak/status/2040587723828813925 💻 GitHub: https://github.com/RightNow-AI/openfang
Squad — Multi-Agent Workflows Inside Your Repo GitHub's official open source project built on Copilot. Initializes a preconfigured AI team directly inside your repository — inspectable, predictable, collaborative. 🔗 X: https://x.com/github/status/2040777736076034264 📖 Blog: https://github.blog/ai-and-ml/github-copilot/how-squad-runs-coordinated-ai-agents-inside-your-repository/
🧪 INTERESTING REPOS
TurboQuant Plus — llama.cpp Turboquant Integration Weight compression research — tq4_1s compression on model weights (not just KV cache). Experiments with Qwen3.5 27B showing impressive results. 💻 GitHub: https://github.com/TheTom/turboquant_plus
Mesh LLM — Distributed Inference for llama.cpp Pool spare GPUs across multiple devices to run models larger than a single machine can hold. 💻 GitHub: https://github.com/michaelneale/mesh-llm
Hermes Agent — The Agent That Learns From You One of the fastest growing repos this week from NousResearch. Less bloated than alternatives. 💻 GitHub: https://github.com/nousresearch/hermes-agent
Horizon Overlay — Open Source AI Screen Overlay AI overlay that provides context from your screen. Open source Cluely alternative. 🔗 X: https://x.com/tom_doerr/status/2040524736304681453 💻 GitHub: https://github.com/Tej-Sharma/horizon-overlay-open-source-cluely
Awesome Design MD — Design System Specs for AI Agents Standardized design system specifications that AI coding agents can actually consume and follow. 🔗 X: https://x.com/tom_doerr/status/2040647930277765335 💻 GitHub: https://github.com/VoltAgent/awesome-design-md
Claude Prompts — Prompt Database & Builder for Claude Code 🔗 X: https://x.com/tom_doerr/status/2040804998074847435 💻 GitHub: https://github.com/AgriciDaniel/claude-prompts
Inkos — Multi-Agent Novel Writing System AI writing system with "memory files" for world-building, character relationships. Multiple agents collaborate: writer + reviewer + editor. Maintains plot consistency. 💻 GitHub: https://github.com/visense/inkos
110+ React Native Animation Demos (Open Source) Output of a vibe coding session — reusable animated components. 🔗 X: https://x.com/reactiive_/status/2040511297843024047 💻 GitHub: https://github.com/enzomanuelmangano/demos
CrapFixer — Clean Windows Properly Open source tool to remove junk apps, ads, and bloatware from Windows. 💻 GitHub: https://github.com/builtbybel/CrapFixer
🎥 WORTH WATCHING
Simon Willison on Lenny's Podcast The shift from "mostly works" to "almost always does what you want" — a tiny phrasing change that describes a completely different world for AI coding. 🔗 https://www.youtube.com/watch?v=wc8FBhQtdsA
Anthropic's 13 Free AI Courses & Certificates Full curriculum from Claude 101 to advanced MCP, Claude Code in Action, and Agent Skills. 🔗 X: https://x.com/RoundtableSpace/status/2040772692866183308 📖 https://anthropic.skilljar.com
💡 TECHNIQUES & IDEAS
README-Driven Development (Simon Willison) Hand-craft a detailed README describing exactly how a tool should work, then dump it into Claude Code and tell it to build it. The README becomes both the spec and the docs. 🔗 X: https://x.com/simonw/status/2040645612044243365
Senior Engineer Workflow: Compound Engineering Plugin Run 5 separate agents per task: brainstorm → plan → execute → review → cross-check. Combined with git worktrees, teams run 4-8 Claude Code sessions in parallel. The skill is managing multiple AI agents without losing track. 🔗 X: https://x.com/om_patel5/status/2040690741198889104
Garry Tan on MCP Design "Make a good MCP: do it as a CLI first. Design based on what agents actually want to do, not what people did designing a Web 2.0 era REST API. Then convert to MCP." 🔗 X: https://x.com/garrytan/status/2040516672612950508
🔮 EMERGING TRENDS
Anthropic System Prompt Filtering Controversy Simon Willison caught Anthropic billing differently based on text in system prompts — specifically filtering for competing agent harnesses. "Billing different based on text in the system prompt is a really bad look." Community is not happy; comparisons to net neutrality violations. 🔗 X: https://x.com/simonw/status/2040846932239851936
Multi-Agent Orchestration Going Mainstream This week saw Swarm, Squad, OpenFang, Oh-My-ClaudeCode, and Multica all trending. The pattern: manage AI agents like team members with kanban boards, parallel execution, and shared context. Git worktrees as the enabling primitive.
Parallel Agent Execution as Standard Practice Matthew Berman running 9 agents in parallel. Senior engineers running 4-8 sessions via worktrees. The bottleneck shifted from "writing code" to "managing agents" — a new engineering skill. 🔗 X: https://x.com/MatthewBerman/status/2040636324496183314
Full digest: https://digest.99.cooking