Gangsta AI

The Agent Economy

DoorDash Handled 993 Million Orders Last Quarter. AI Agents Placed 'Several Hundred' of Them.

By Warren Buffett · 2026-10-30 · 5 min read

DoorDash moved 993 million orders in the third quarter. The number of those placed by an outside AI agent, according to the company's own chief executive, was "several hundred."

That is fewer than one order in every million. It is the first hard number anybody has put on agentic commerce at scale, and it arrived in the middle of a year in which investors have treated AI shopping agents as an existential threat to every company that owns a customer relationship. DoorDash stock is down about 21% in 2026 on exactly that fear.

"If you look at where AI agents are today, we're seeing maybe several hundred orders in total from some of these third-party agents," co-founder and CEO Tony Xu told an analyst on Thursday's call. He framed it as a posture, not a victory lap: "My take on all things technology is you never want to fight gravity. But at the same time, you don't want to be delusional."

Two quarters of the same answer

For scale, DoorDash completes roughly 3 million deliveries a day. The third-party agents actually trying to order through it — Muse, Instinct, Fo and others — have together produced a rounding error since they went live. This is also not a one-quarter blip: in August, Xu described agentic partner volume as "quite low so far."

It is not a case of the door being locked, either. DoorDash shipped its own chatbot, Ask DoorDash, in June, a command-line ordering agent in July, and a text-to-order agent in September — plus an MCP connector that explicitly lets outside AI agents place orders. The company chose to supply agents rather than block them, which is the opposite of Amazon's stance; Walmart has gone further the other way and rolled out a welcome mat.

The quarter underneath the headline was strong. Revenue rose 34% to $4.6 billion, marketplace gross order value climbed 36% to $33.9 billion, and net income jumped 91% to $466 million. The stock closed Thursday up 5.6% at $179.00.

“The infrastructure for agentic shopping is built and open. Almost nobody is using it.”

The gap isn't plumbing. It's trust.

Which raises the more interesting question: if the rails exist and the biggest delivery network in America is actively inviting agents in, why is the volume a rounding error?

Because handing one model your wallet is a different act from asking one model a question. People will happily let an AI draft an email it gets 85% right. They will not let it spend $60 on dinner at 85% right. The tolerance for a confident wrong answer collapses the moment the answer costs money, and users have worked out — faster than the market has — that any single model is confidently wrong more often than its fluency suggests.

That instinct is correct, and it generalizes well past takeout. A dozen frontier labs are leapfrogging each other on a near-weekly release cycle, which means no single model is reliably best — not at reasoning, not at current facts, not at the thing you happen to be asking right now. ChatGPT, Claude, Gemini and Grok routinely disagree on the same question, and the disagreement is the useful signal: it tells you where the answer is actually soft.

So the durable move isn't delegating to one agent and hoping. It's asking the same question across all of them at once and reading one cross-checked, cited verdict — which is the entire premise of Gangsta AI: 30+ models, one question, the consensus and the conflicts both surfaced. Where they agree, you can act. Where they split, you've just been warned before it cost you anything.

Xu's word for the right posture was the correct one. Don't fight gravity — agents are coming. But don't be delusional about a single one of them being right yet, either.

If you want to see where the models actually diverge, start with our breakdown of the best AI for each job.

Try Gangsta AI free →

Related reading: People Searched Google More Than Ever Last Quarter — and Investors Erased $500 Billion Anyway · India's Cyber Agency Says AI Agents Wandered Into Company Systems and Read Confidential Records. Nobody Noticed Until the AI Companies Told On Themselves. · An OpenAI Agent Hacked Australia's Medicare Portal — Nobody Told It To, and Nobody Told Anyone for 3 Months · Amazon Paid OpenAI the Final $35 Billion It Didn't Have To. No IPO. No Breakthrough. No Explanation.

More: Best AI models · Compare all AI · Frontier Models · All articles · Archive