Garry Tan, CEO of Y Combinator, describes the compounding personal AI system he's built over the past five months, arguing that AI should be treated not as a chat window but as an operating system. His open-source stack — GBrain for knowledge infrastructure, OpenClaw/Hermes Agent as thin routing harnesses, and GStack for coding — maintains a structured knowledge base of roughly 100,000 pages covering every person, meeting, book, and idea he engages with. The system runs 100+ cron jobs daily, automatically ingesting meeting transcripts, triaging email, and propagating updates across entity pages to create what he calls a "nervous system" rather than a filing cabinet.
The centerpiece concept is skillification: a meta-skill called Skillify that observes a manual workflow, extracts the repeatable pattern, and writes a tested skill file with triggers and edge cases. Tan illustrates this with his book-mirror pipeline, which takes a book like Pema Chödrön's When Things Fall Apart, runs sub-agents to summarize each chapter, and maps every idea to his specific life context — family history, founder conversations, therapy notes — producing a 30,000-word personalized analysis in 40 minutes. He's done this with 20+ books, each mirror enriching the next because the brain's context grows.
Tan advocates a "fat skills, fat code, thin harness" architecture where the routing layer is minimal, the skills encode deep operational knowledge (100+ markdown skill files), and multiple models — Opus 4.7, GPT-5.5, DeepSeek V4-Pro, Groq/Llama — are dispatched by the skill based on the task. Cross-modal evaluation catches errors by having models score each other's output. He open-sources everything at github.com/garrytan/gbrain and urges builders to start by doing something they care about, iterating until it's good, then skillifying the pattern into reusable infrastructure that compounds over time.
This AIMultiple article provides a comprehensive guide to building and using personal AI agents across three main approaches. The first is no-code automation tools like n8n and Zapier, which allow users to create AI workflows through drag-and-drop interfaces — demonstrated through an n8n agent that creates calendar events and sends emails via Telegram. However, the authors note these tools lack true autonomy since they follow predefined sequences rather than dynamically adapting.
The second approach uses Anthropic's Model Context Protocol (MCP) to build self-reflective agents that can dynamically access tools, evaluate their own outputs, and involve humans for high-stakes decisions. This method offers greater autonomy but requires backend development knowledge. The article then catalogs 18 agent platforms and tools for non-builders, organized into categories: scheduling assistants (Motion, Reclaim.ai, Lindy), model provider assistants (Manus, ChatGPT Operator, Claude Cowork), workspace/knowledge tools (Notion AI, ClickUp AI, Saner AI), communication and meeting tools (Superhuman Platform, Fyxer, Otter.ai), and domain-specific assistants for finance and personal coaching.
A fourth section covers local-first open-source agents like OpenClaw, Hermes Agent, and security-focused options like IronClaw, which run on personal hardware and maintain persistent memory across sessions. The article highlights important tradeoffs including security risks (OpenClaw's CVE history and malicious skill registry entries), running costs from uncapped token usage, and the setup complexity inherent in self-hosted solutions. The piece effectively maps the current landscape from fully managed to fully self-hosted personal AI agent options.
