SkillSpector is NVIDIA's security scanner for AI agent skills. It vets a skill before install to find prompt injection, data exfiltration, and malicious code. We analyze its structure — a LangGraph map-reduce graph that fans out to 25 analyzers (static patterns, AST, taint, YARA, MCP, LLM semantics) and reduces them to a single risk score — against Superpowers and ponytail, which inject skills.
ponytail packs a single discipline — the "lazy senior dev" laziness ladder — into one SKILL.md and ships it to 16 agents via every mechanism (skill, hook, command, MCP, plugin), proving a ~54% code cut with a real agentic benchmark. We analyze it against Superpowers, and the idea that a skill is a bundle of discipline, distribution, and measurement rather than code.
Browser Use is a Python agent in which an LLM drives a real browser. It pre-chews the page into an indexed list of interactive elements for the LLM, drives the browser via CDP instead of Playwright, and runs the session with an event bus and watchdogs. We look at why this fits LLMs better, picking up from the earlier Playwright analysis.
Cline is a coding agent that lets you approve every action inside VS Code and roll back at any time. It also detaches its core from the host via a protobuf/gRPC boundary (ProtoBus + HostBridge), so the same core runs in the editor and in a terminal CLI. We analyze Plan/Act approval, checkpoints, and the host abstraction, contrasting them with OpenCode's headless engine.
OpenCode externalizes model metadata to models.dev and even hand-rolls its own LLM protocol layer, so any provider attaches with a single line of data. A single Effect-based HTTP engine is shared by the TUI, web, desktop, Slack, and editors (ACP), while two generations — legacy and V2 — coexist. We analyze the structure against Qwen Code's single-vendor platform.
A retrospective on building a macOS desktop app (Tauri + Rust + React) with an AI agent. The decision to make markdown files — not the database — the source of truth, the research I did before writing any code, the features I chose not to build, and the macOS permission bug that took the longest to crack. A story about judgment more than code.
Semble is a Python library that splits code into chunks with tree-sitter, fuses Model2Vec static embeddings with BM25 via RRF, and applies code-aware reranking — delivering millisecond code search on CPU alone. Where CodeGraph solves the same problem with an AST knowledge graph, Semble solves it through retrieval. This post analyzes the architecture by contrasting the two approaches.
CodeGraph is a TypeScript tool that uses tree-sitter to parse source code, builds a local SQLite knowledge graph of symbols, edges, and files (with FTS5), and exposes that graph via an MCP server to coding agents like Claude Code, Cursor, Codex, and Hermes Agent. Instead of burning tokens navigating a codebase with grep/Read calls, agents get answers from a single codegraph_explore invocation. This post analyzes the architecture that makes that possible.
GitHub's spec-kit and Anthropic's superpowers plugin both force a workflow onto coding agents so they never drift into vibe coding. But one is a spec-first file system that leaves the specification behind as an artifact in a .specify/ directory, while the other is a collection of discipline prompts lazily loaded through the Skill tool. We compare the two projects across distribution model, artifact philosophy, token cost, and extensibility.
WeKnora is a Go-based enterprise knowledge framework open-sourced by Tencent. It bundles document parsing, vectorization, hybrid search, and LLM inference into an event-driven chat pipeline, then layers a ReAct Agent and Wiki Mode on top. This analysis covers how a Python docreader gRPC service, 20+ LLM providers, 7 vector DBs, 7 IM channels, multi-tenant RBAC, and Langfuse observability are all handled as swappable components within a single monorepo.
Dify is an open-source project that brings LLM app development, a workflow canvas, a RAG pipeline, model/tool plugins, MCP, and operational observability together into a single productized platform. This post analyzes its architecture through the lens of the Flask API, Graphon workflow runtime, Celery workers, Next.js console, plugin daemon, and vector backend structure.
OpenHands productizes an AI coding agent across a local GUI, FastAPI app server, sandbox, agent-server, SDK, event store, MCP, and skill system. This analysis focuses on the boundary between the app server and GUI that the current OpenHands repository is responsible for.
Qwen Code is a TypeScript-based coding agent that bundles a terminal CLI, an LLM provider abstraction, a tool scheduler, MCP, Skills, Subagents, the qwen serve daemon, channel plugins, and IDE integration into a single repository. We analyze how it reconstructs a Claude Code-style experience on top of Qwen/DashScope, multiple providers, and an extensible agent runtime.
Ruflo extends Claude Code with a CLI, MCP server, swarm coordination, AgentDB memory, hooks, background workers, a plugin marketplace, and a Web UI to create a multi-agent operations layer. This post analyzes the architecture of Ruflo — an evolution from Claude Flow — and connects it to earlier posts on agentmemory, Superpowers, and Hermes Agent.
Keeping Claude Code, Codex, Gemini CLI, and OpenClaw in their own separate silos has clear limits. agentmemory collects observations via hooks, reconstructs them with BM25, vector, and graph retrieval, and creates a long-term memory layer shared across multiple agents through MCP, REST, and a viewer.
An architecture deep-dive into Nous Research's Hermes Agent from a user perspective. Covers the CLI, messenger gateway, ACP, tool registry, skills, memory, plugins, and sub-agent structure in an accessible, flow-oriented way.
An analysis of how Claude Code Game Studios combines 49 agents, 72 skills, hooks, rules, and templates to run a single Claude Code session like a game development studio.
A practical guide to integrating RedisBloom into a signup duplicate-check flow, covering request routing, sharding, synchronization, rebuild strategies, and monitoring.
A deep dive into the Playwright repository — covering the core engine, client/server protocol, fixture-based test runner, why it has become so popular, why it feels slow when paired with an LLM, and what the alternatives look like.
An architectural analysis of Paperclip, an open-source control plane that operates AI agents as a single virtual company under an org chart, budget, and governance structure.
An analysis of Steve Yegge's Beads project. A deep dive into its Dolt-backed distributed graph issue tracker architecture, designed to give AI agents structured memory, dependency management, and the ability to execute long-horizon tasks.
A deep-dive into the architecture of agent-browser, Vercel Labs' Rust-based browser automation CLI for AI agents — covering CDP-based control, the accessibility-tree Ref system, Provider abstraction, and the security model.
A comprehensive architecture analysis of Ollama, a local LLM execution platform implemented in Go. Covers the tech stack, scheduler, model management, llama.cpp integration, and the Runner system in detail.
A comprehensive architectural analysis of LangChain, the Python-based LLM application framework. Covers the full technology stack, core modules, the Runnable interface, the callback system, and partner package structure in detail.
A comprehensive architecture analysis of NanoClaw, a TypeScript-based Personal Claude Assistant. Covers the single-process orchestrator, container-isolated agents, channel skill system, and security design in detail.
A deep-dive into the architecture of Lightpanda, a Zig-based ultra-lightweight headless browser. Covers the technology stack, DOM engine, V8 JavaScript integration, CDP protocol, and network layer behind its 9× lower memory usage and 11× faster speed compared to Chrome.
A comprehensive architecture analysis of OpenClaw, a TypeScript-based Personal AI Assistant framework. Covers the technology stack, core modules, message pipeline, and channel system in detail.