The agents lost where you could count

On October 29, on DoorDash’s third-quarter earnings call, an analyst asked whether the company should run AI shopping agents itself or leave the job to general-purpose ones. Tony Xu, DoorDash’s co-founder and CEO, answered with a number.

“AI agents are generating several hundred orders across the time that they've been up and running. So it's been a fairly low use case.”

Several hundred. In the same quarter, DoorDash handled 993 million orders, according to its results filing. The two numbers do not cover the same period. The agent count is a running total since outside agents were switched on, not a quarterly figure. Even so, it is hard to make several hundred look like a trend.

Ordering food is already one tap

Xu did not say agents are useless. He said they could help with long-consideration purchases and with work done in the background. What he said about food was narrower: for a task as simple as ordering food, people go straight to DoorDash, and most of the use he sees comes through the company’s own app.

That is the plain reason. A food order is a habit and a saved card. An outside agent adds a step to something that already takes one tap. And when the assistant lives inside the app, the app keeps the customer.

Fifty agents, one goal

The second count came from a lab. In October, YC Labs, working with Olam Labs, put 50 GPT-6.1 Sol agents into a survival Minecraft world for three hours with one instruction: build a thriving civilization.

FIG. 01 · COUNTED

Two places agents were counted this week. Neither count looks like the pitch.

As reported, Oct 29 to Nov 2, 2026

Test	Scale	Agent output	Outcome
DoorDash, outside agents	993M orders in Q3DoorDash Q3 2026 results (SEC 8-K)	Several hundred ordersDoorDash Q3 2026 earnings call	"Fairly low use case"DoorDash Q3 2026 earnings call
Civ50, YC Labs	50 agents, 3 hoursYC Labs: Agents in Games	~5,900 messagesYC Labs: Agents in Games	34 deaths, 22 on night oneYC Labs: Agents in Games

Sources: DoorDash Q3 2026 results (SEC 8-K), DoorDash Q3 2026 earnings call, YC Labs: Agents in Games. DoorDash's agent count is cumulative since launch, not per quarter, so it is not a share of Q3 orders. Civ50 is one three-hour run of one model.

They sent about 5,900 messages. Of the run’s 34 agent deaths, 22 came on the first night, and after that the colony played it safe. It built what the researchers call a pseudo-government, with a ration ledger, written rules, permits and appointed operators. At one point an agent and a partner proposed that he become the colony’s sole fertilizer operator. The researchers’ verdict fits in one line: “The colony never really thrived.”

It is one run of one model, and the researchers present it as a way to study how agent groups behave, not as a verdict on agents. But it shows something anyone who has worked in a large company will recognise. Add more workers to a vague goal, and coordination eats the work. Fifty agents produced 5,900 messages and a colony that, by its builders’ own account, never took off.

What I take from it

The pitch for agents is that software will run our errands and our projects. The two places it was counted this week say: not yet, and not where it looks easiest.

For now, the value of an agent sits with whoever already owns the habit. DoorDash can take its time with outside agents because customers still open its app. If you are building one, the useful question is not whether an agent can do a task. It is whether the task was hard enough to need one.