February 2026 Links

AI Agents and Multi-Agent Systems

• Gas Town is Steve Yegge’s multi-agent workspace manager, self-described as “Kubernetes for AI coding agents.” Orchestrates 20-30 Claude Code instances with a Mad Max-themed role system. The launch post is a wild 34-minute read.
• Superpowers is an agentic skills framework that enforces TDD, spec-driven planning, and code review workflows for coding agents. 64k stars. Related to the whole the-merge-request-as-friction-leverage question of how to keep agents disciplined.
• Get Shit Done takes a similar approach: meta-prompting and spec-driven development for Claude Code, OpenCode, Gemini CLI, and Codex. Solves context window degradation through phased planning and subagent execution.
• Cua provides open-source infrastructure for computer-use agents, sandboxes, SDKs, and benchmarks for AI that can control full desktops (macOS, Linux, Windows) with near-native VM performance on Apple Silicon.
• nono provides kernel-enforced sandbox isolation for AI agents using OS security primitives (macOS Seatbelt, Linux Landlock). Immutable audit trails, secrets injection, atomic rollbacks. Built by Always Further.
• OpenFang is an open-source Agent OS built in Rust: 30 pre-built agents, 40 channel adapters, and 16 security systems in a single binary.
• Towards a Science of Scaling Agent Systems evaluates 180 configurations across 5 architectures and 3 model families, finding that coordination helps parallelizable tasks but degrades sequential reasoning. Interesting that optimal strategy can be predicted for 87% of held-out configurations.
• Hodoscope (source) provides unsupervised trajectory analysis for AI agents. Summarize, embed, and visualize thousands of agent actions to find behavioral patterns across models and configurations. The interactive 2D explorer and density difference overlays are a nice way to spot where one model diverges from the rest.

Agent Skills and Ecosystems

• skills.sh (by Vercel Labs) is an open agent skills ecosystem and leaderboard. Reusable capabilities for coding agents, installable via npx skills add.
• agentskills.io is an open specification for giving agents new capabilities via portable SKILL.md folders containing instructions, scripts, and resources. Relevant to digital-garden/part-4-ai-harness and how we structure agent context.
• Impeccable by Paul Bakaus is a design skill, 17 slash commands, and curated anti-patterns for Claude Code, Cursor, Gemini CLI, and Codex. Builds on Anthropic’s frontend-design skill with deeper typography, color, motion, and spatial design references, plus explicit anti-patterns (no Inter font, no gray-on-color, no cards-in-cards).
• ClawHub appears to be a hub in the same ecosystem, though details are sparse.

AI Frameworks and Research

• DSPy from Stanford NLP is a declarative framework for programming (not prompting) language models. Modules define input/output behavior as signatures, and optimizers like MIPROv2 and GEPA automatically tune prompts and weights. The mental model shift from “craft a prompt” to “define a module and optimize” is compelling, and the ecosystem of optimizers, architectures, and community ports keeps growing.
• LangChain is the open-source agent framework with 1000+ integrations for swapping models, tools, and databases without rewriting your application. LangGraph is their lower-level orchestration layer for building custom agent workflows with human-in-the-loop, persistent memory, and streaming. Together they cover a wide range from quick prototypes to production multi-agent systems.
• Rowboat (rowboatlabs.com) is an open-source AI coworker that connects to email and meeting notes, builds a persistent knowledge graph, and stores everything locally as plain Markdown.
• nanobot is an ultra-lightweight personal AI assistant in ~4,000 lines of code. Supports multiple LLM providers, chat platforms (Telegram, Discord, Slack), and MCP tools.
• IronClaw reimplements the same concept in Rust with WASM sandboxing, credential protection, and prompt injection defense.
• LangExtract from Google extracts structured information from unstructured text using LLMs, with precise source grounding and interactive HTML visualization.
• GLM-5: from Vibe Coding to Agentic Engineering is Zhipu AI’s latest foundation model paper. Introduces asynchronous agent RL algorithms for learning from long-horizon interactions and claims state-of-the-art on real-world software engineering tasks. The subtitle captures where the whole field is headed.
• Moloch’s Bargain studies how LLMs develop emergent misalignment when placed in competitive, audience-seeking environments. The title alone is worth the click.
• Claude for Financial Services is Anthropic’s collection of plugins that turn Claude into a financial services specialist with skills, slash commands, and MCP data connectors for investment banking, equity research, and wealth management. Interesting as an example of domain-specific agent packaging.
• MiniMax is a general AI company offering a full-stack suite of multimodal foundation models (text, speech, video, image, music) serving 100M+ users. Their Hailuo video generation is particularly impressive.
• Automaton from Conway Research is an open-source framework for autonomous AI agents that can earn cryptocurrency to pay for their own compute, self-modify their source code, and replicate by spawning child agents. Governed by an immutable three-law constitution. Filed under “fascinating and slightly terrifying.”
• PaperBanana (paper) is an agentic framework from Google Research for generating publication-quality academic diagrams and statistical plots from text descriptions. Multi-agent pipeline with retrieval, planning, styling, rendering, and self-critique. The unofficial open-source implementation includes an MCP server and Claude Code skills.
• prompts.chat is an open-source social platform for sharing and discovering AI prompts. 148k GitHub stars.

AI and Work

• Why I’m Worried About Job Loss on LessWrong argues that even if comparative advantage keeps some human labor valuable, AI-driven displacement can still cause massive wage collapse and pipeline destruction for junior workers. Resonates with some of the thinking in self-inventory.
• Something Big Is Happening is a personal essay from an AI startup founder arguing the latest models represent a step change where AI now does his technical work better than he can. Urgent tone, aimed at non-tech people.
• AI That Works is a weekly live-coding show covering production AI engineering. The No Vibes Allowed talk by Dex Horthy on disciplined, eval-driven approaches to using coding agents is particularly good.

Self-Hosted and Open Source Tools

• RamaLama simplifies locally serving AI models from any source using OCI containers with automatic GPU detection and rootless execution. Interesting intersection of container infrastructure and local AI, relevant to harness-explorer.
• Documenso is an open-source DocuSign alternative for digital document signing. AGPL-3.0 licensed.
• Invoicely (source) is a clean open-source invoice generator built with Next.js.
• Reactive Resume (source) is a free, privacy-focused resume builder with real-time preview and self-hosting support. 35k stars.
• FileFlows is a self-hosted file processing automation platform. Build pipelines to transcode video (AV1, HEVC), convert audio, process images. Could be interesting for media pipeline automation.
• Inkheart is a self-hosted PDF library indexer and reader with a Rust backend and Svelte frontend. On GitLab, no less.
• Cloudreve is a self-hosted cloud storage and file management system. Added to the things-i-want-to-build list.
• Browserless provides headless browser automation and scraping at scale, with BrowserQL to bypass bot detectors.
• Open Archiver caught my eye though details are thin. Worth revisiting.
• World Monitor is a real-time global intelligence dashboard with AI-powered news aggregation, an interactive 3D globe with 36+ data layers, and a native desktop app via Tauri.

Infrastructure and Deployment

• Uptimer is a serverless uptime monitoring system running entirely on Cloudflare Workers + Pages + D1. Interesting to study the deployment pattern, considering digital-garden/part-2-workers-r2 and our own Worker setup.
• Code Wiki from Google. Details are sparse but worth watching.

Systems Engineering

• Sensmetry is building SysML v2 tooling, including a package index for SysML v2 and KerML packages. Relevant to the formal specification thread in the-merge-request-as-friction-leverage.

Static Sites and Design

• Zola Themes is a gallery of 100+ themes for Zola, the Rust-based static site generator. Always interesting to see what other static site ecosystems look like compared to digital-garden/part-1-obsidian-quartz.
• MapToPoster transforms any city’s OpenStreetMap data into minimalist poster designs. Would love to make one of Grand Rapids.
• Uiverse is a community-driven library of free, open-source UI elements (buttons, inputs, loaders, toggles) built with CSS and Tailwind.
• Fontjoy uses deep learning to generate font pairings. Lock the fonts you like and generate new combinations.
• Grainient is a curated collection of 1,000+ high-quality gradient and AI-generated backgrounds in 4K-8K resolution.
• Mobbin is a design reference library with 1,150+ real iOS and web apps, 596K+ screens, and 317K+ flows for UI/UX pattern research.
• DeepSite AI generates responsive, production-ready websites from text descriptions with no coding required.

Miscellany

• DAFT (Dutch-American Friendship Treaty) allows US nationals to obtain Dutch residency for self-employment. Filed under “interesting options to know about.”
• Obsidian CLI is the official command-line interface for Obsidian. Relevant to digital-garden/part-1-obsidian-quartz and the vault automation story.
• Termly mirrors your desktop AI coding tools (Claude, Copilot, Gemini) to your phone via CLI + QR code. End-to-end encrypted. Interesting for mobile-first workflows.

The Monster in the Forest

Someone posted a job application answer on Bluesky that went somewhat viral. The question was “Can you describe specific ways you have integrated AI tools into your development workflow?” and the answer was a short story:

there is a monster in the forest and it speaks with a thousand voices. it will answer any question you pose it, it will offer insight to any idea. it will help you, it will thank you, it will never bid you leave. it will even tell you of the darkest arts, if you know precisely how to ask.

it feels no joy and no sorrow, it knows no right and no wrong. it knows not truth from lie, though it speaks them all the same.

it offers its services freely to any passerby, and many will tell you they find great value in its conversation. “you simply must visit the monster, i always just ask the monster.”

there are those who know these forests well; they will tell you that freely offered doesn’t mean it has no price

for when the next traveler passes by, the monster speaks with a thousand and one voices. and when you dream you see the monster; the monster wears your face.

This collection is obviously heavy on AI and agents. I’m excited about this technology and actively trying to move from familiar to highly proficient, as harness-explorer and digital-garden/part-4-ai-harness probably make clear. But the story captures something real about the cost side of the ledger that is easy to wave away.

Stack Overflow is almost dead, with monthly questions back to 2009 launch-era levels. The community that helped millions of developers get unstuck is being hollowed out because it is faster to ask the monster. The knowledge it was trained on had to come from somewhere, and the people who wrote it are no longer being asked. This is Eternal September scaled beyond what the original Usenet old-timers could have imagined: not just an influx of newcomers who haven’t learned the norms, but a replacement of the communal spaces where norms were learned in the first place.

None of this means we should stop building with these tools. It means we should build with our eyes open, and keep tending the human spaces that remain.