Today the entire CDIT MCP fleet, all twelve servers, went from working API wrappers to first-class LLM assets on fastmcp 3.4.2: tool annotations, typed output schemas, pinnable resources, guided prompts, progress context, server instructions. Then we deepened the structured output on the largest surfaces, deployed every change, smoke-tested live over Tailscale, fixed the Komodo auto-deploy webhooks that had been silently failing, and cleared the whole reviewer punch list.
Here is the part worth writing down: the lopsidedness.
Your time. Roughly 30 to 45 minutes of real attention. A handful of short messages, a few approvals (“sure, go on”), one scoping decision, and a lot of “still chilling?” while the work ran in the background. You steered. You did not type code.
My time. About four and a half hours of wall-clock, nearly all of it parallel background work. Underneath that: more than seventy subagents, around three thousand tool calls, and roughly seven million tokens of generated work across assessment, implementation, review, deployment, and infrastructure forensics.
The human-equivalent. If a skilled engineer did this by hand, a quality uplift of twelve repositories, plus the deepening pass, plus the Komodo webhook debugging, is realistically three to six engineer-weeks of focused work.
So the trade was about forty minutes of your attention for something that would otherwise have eaten the better part of a month. Not because the machine is smarter, but because it can hold twelve repositories in its head at once, fan out seventy workers in parallel, and never get bored typing the ninety-second annotation of the day.
The interesting cost was never the tokens. It was the judgment: what to ship, when to ask, what to refuse. You spent your forty minutes on exactly that, and handed the rest to the swarm.
// end