Agent upskilling trend
Builders are stacking custom agent configs (CLAUDE.md‑style files) and orchestration to 10x output — some creators report 50+ custom skills as a real moat over generic prompts. Educators and practitioners are already talking about an "Udemy for agent upskilling," teaching how to break projects into 15–20 minute agent tasks and manage agent memory limits like 'goldfish' working memory. (x.com) (x.com)
AGENTS.md and CLAUDE.md formats now appear in large-scale project tooling: one guide cites AGENTS.md as present in 60,000+ Git repos. (claudelab.net) A vendor-level feature set for packaged skills is live: Anthropic’s Agent Skills are documented as modular SKILL packages in Claude, and OpenAI’s Codex docs specify a SKILL.md folder format that agents load from repo and system locations. (platform.claude.com (developers.openai.com) Community curation is extensive: a GitHub collection lists 1,000+ agent skills and has 13.4k stars, while a separate “awesome” repo for agent skills has 3.4k stars and active commits. (github.com/VoltAgent/awesome-agent-skills (github.com/heilcheng/awesome-agent-skills) Dedicated marketplaces and registries are scaling: SkillsMP advertises a catalog of hundreds of thousands of downloadable agent skills, and regional hubs such as SkillHub report curated top-50 skill lists. (skillsmp.com (skillhub.tencent.com) Commercial training has responded: Udemy’s “Deep Agent” course runs about 20 hours across 183 lectures and teaches multi-agent orchestration, and bootcamps on LangGraph/LangChain advertise hands-on agent state, memory, and orchestration modules. (udemy.com/course/deep-agent (classcentral.com/course/udemy-complete-agentic-ai-bootcamp-with-langgraph-and-langchain-458828) Cloud vendors document orchestration and custom orchestrators: Microsoft published AI agent design patterns for sequential and handoff orchestration last month, and AWS Bedrock provides a “custom orchestrator” example for fine-grained workflow control. (learn.microsoft.com (aws-samples.github.io) Memory and agent architecture are now course material: recent explainers and videos describe multi-component memory (working, semantic, episodic, procedural) and argue for small working-memory windows to reduce hallucination, and practitioner posts document how skills inject instructions without bloating context. (youtube.com/watch?v=wEa6eqtG7sQ (tylerfolkman.substack.com) Practical evidence of scale exists in config collections: a public repo of Claude Code configs describes “production-ready agents” developed over 10+ months, supporting the emergence of dozens-to-hundreds-of-skill libraries rather than one-off prompts. (github.com/eddiemessiah/config-claude-code)
