In December 2025, 📝Andrej Karpathy — one of the most accomplished AI researchers alive — publicly said he had never felt so behind.
The post named a phase shift in programming itself. The bits a programmer contributes are increasingly sparse. The leverage now lives in how cleanly someone strings together a stack of agentic tooling — agents, subagents, prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations, and the mental models that hold the whole apparatus together. The discipline is no longer writing code; it is composing the conditions under which agents write it well.
The convergence is what gives the post its weight. Karpathy's enumeration mirrors what Anthropic has been formalizing under the banner of context engineering, what the Claude Code ecosystem has been building piecewise through hooks, skills, slash commands, and subagents (see 📝AGENTS.md (Claude Code)), and what the broader MCP movement has been wiring up between models and external systems. The same shape shows up in the 📝agentic engineering framing he would publish two months later. Independent threads, same architecture.
The application is not to learn each item on the list. It is to notice that the surface area of leverage has moved from the editor to the orchestration layer — and that the people compounding fastest are the ones building durable substrates underneath their agents, not the ones with the cleverest single prompts.