2025: The Year Agents Went to Work

A year ago this week I wrote about o3 breaking the ARC benchmark and said 2025 would be the year agents left the lab. Grading that call, and the year, honestly:

What changed — capability. January's Operator fumbled through webforms. By July, ChatGPT ran a virtual computer; by autumn, coding agents worked for hours unsupervised, and every major lab shipped a frontier model in the fourth quarter alone — GPT-5.1, Gemini 3, a new Opus. The capability story never paused for breath.

What changed — adoption. The default flipped. Shopify's memo in April ("prove AI can't do it before hiring"), agent modes in every enterprise suite by summer, and by year-end, layoff letters that say the machine's name. Weekly ChatGPT usage approaching a billion people. "Do you use AI at work?" became as dated a question as "do you use the internet?"

What changed — evidence. This was the year the labor data arrived: Stanford's canaries showing a 13% relative employment decline for the youngest workers in exposed occupations; new-grad unemployment inverting above the national average; Anthropic's own index showing automation-style usage climbing. The debate moved from "will it affect jobs?" to "how fast, and for whom?" — which is exactly the debate I asked for in 2024.

What didn't change. Policy. Another full year of the most important labor transition in a century, and no major economy stood up an income-floor pilot at scale, a transparency requirement on automation separations, or a serious apprenticeship program for the AI era. The vacuum where the response should be is now the single biggest risk factor I track — bigger than any capability threshold, because capability is inevitable and response is a choice.

For 2026

Watch three things. Agents with real autonomy budgets — spending, hiring, committing their principals. Humanoids leaving the demo reel for actual pilot deployments at scale. And the first political campaigns anywhere that run on the transition rather than around it. I intend to spend the year the way I spent this one — documenting, warning, and building the playbook I keep asking institutions to write.

The machines clocked in this year. The question of what the rest of us are paid to do — and who decides — is now on the clock too.

LLAP. 🖖