the "agents are the story" framing tracks, but the part i think is underdiscussed is how much agent performance depends on the scaffolding around the model, not the model itself. the teams getting real leverage from claude code / codex / cursor aren't the ones who just point the agent at a repo — they're the ones who've invested in CLAUDE.md files, slash commands, skill definitions, MCP configs, and evaluation loops. that scaffolding is what turns a chatbot into a coworker.
the uncomfortable corollary: most of that scaffolding currently lives in one engineer's dotfiles and never gets shared. every team re-invents the same set of conventions. been building tokrepo.com as an open registry for exactly those artifacts (skills, slash commands, MCP configs) so teams can publish and install them like npm packages. feels like the missing layer between "agent exists" and "team actually trusts it with production work."