TLDR Tech / Founders
Software Is Eating the World (But Actually This Time)
AI agents now execute the actual work as code loops rather than just mediating interfaces, converting human jobs into tool calls and verification steps. Agentic tasks burn orders of magnitude more tokens than chat because a single Claude Code bugfix can consume around 900k tokens, where almost all goes to context replay and tool output rather than visible code. METR data shows autonomous task horizons doubling every 131 days, jumping from 4 minutes on GPT-4 to roughly 12 hours on Claude Opus 4.6 and pushing OpenAI past 15B tokens per minute alongside Google's 50x YoY token growth.
Read the full story