The Humans Quietly Running the Agent Economy

Duckbill sells an AI assistant that runs your errands, and this year it bolted on a second kind of worker: an API endpoint that lets any developer hand a task to a vetted human whenever the model can't finish it. The list of tasks worth routing is short and specific — a phone call that has to sound human, a scheduling knot, a CAPTCHA, the little test built to keep machines out ^[1]. The pitch came in familiar terms.

We work on a model that is not dissimilar from Uber, where we boil tasks down into atomic bite-sized pieces. — Meghan Joyce

The same purchase turns up at every layer of the agent business, and no one treats it as a confession. Dazzle AI's chief executive, Marissa Mayer, puts it in the sales pitch.

We're not even trying to hide it, we think the human in the loop is a benefit. — Marissa Mayer

What each layer buys its human for is the question that runs down the whole stack. Start with the labs that are selling the future. OpenAI is hiring its own forward-deployed engineers to sit inside client companies and set the guardrails, on the stated view that the bottleneck is no longer model capability but the human work of getting the thing to run in production ^[2]. Then the vendors who sell the replacement to everyone else. In February, Workday's chief executive Aneel Bhusri asked investors the question nobody selling this stuff says out loud — what agents could do to replace human labor, and what to do with the people who'd be displaced ^[3]. By the end of the day a spokesperson had recast the remarks as comments on general industry shifts. The stock fell 10 percent on slow AI-revenue projections ^[3]. In October the same company reported AI recurring revenue approaching $600 million, with about 900 customers buying agents, in the same release as a 2.5 percent workforce cut ^[4]. The buyers, meanwhile, are bolting the human back on. Atlassian built the hedge into the protocol itself: every agent gets its own profile and an owner attribution, so no one gets to point at the machine ^[5]. Bank of America runs a human-in-the-lead model where the AI does the analysis and people keep the judgment ^[6], and its academy head, Ike Hampton, put the boundary in plain words.

Agents don’t own outcomes. — Ike Hampton

Workday's finance chief, Zane Rowe, has redefined what success even means — what gets measured now is better decisions and more organizational capacity, not return on investment or raw output ^[7]. A metric that rewards judgment instead of throughput has a person hidden somewhere inside it. Then the lenders. Wall Street is still writing the checks, but the terms have tightened: smaller data-center operators have been effectively shut out of the high-yield market ^[8]. And the industry financing the build-out is hedging inside its own walls — Bank of America is training all 211,000 of its people to hold the judgment the agents cannot ^[6]. At the bottom of the stack, the human isn't a hedge. It's the work itself. Lucy Guo, who cofounded Scale AI, describes the new workweek plainly.

I think AI is making people work harder actually because you have to be awake for your agents to be running. — Lucy Guo

The job has acquired a name — botsitting — and a set of symptoms, brain fry, from watching agents closely enough to catch their errors ^[9][10]. The agents run on a consumption model — model calls, data retrieval, API connections, human review, all charged as it's used ^[11]. The initial budgets don't carry this. Most of this work goes unnamed, hidden the way AI use itself is hidden, because workers have learned that disclosure makes them look lazy — 57 percent of employees use AI without telling anyone ^[12]. Then there are the people selling the headline, and their own walk-backs are the last exhibit. Anthropic's two founders cannot agree on the record ^[13]. Dario Amodei can already see where this ends.

I can see it within Anthropic, where I can look forward to a time where on the more junior end and then on the more intermediate end we actually need less and not more people. — Dario Amodei

His cofounder, Daniela Amodei, puts the same future in the opposite frame.

I continue to believe that humans plus AI together actually create more meaningful work, more challenging work, more interesting work, high-productivity jobs. — Daniela Amodei

Bhusri asked the replacement question in February and walked it back the same day; by October he had retreated into the platitude.

Great companies are not built on one generation of technology. They pass the test of time. — Aneel Bhusri

The word replacement keeps getting said and then unsaid, sometimes within the day. What is actually for sale, so far, is not a machine that replaces people. It's the seam between tasks — the part somebody has to hold while awake at odd hours, accountable when the agent fails, carrying judgment no one put a price on by name. The people holding it are the ones no budget line covers, and no one has gotten around to naming.