DoorDash: barely "a few hundred" orders placed by AI agents

DoorDash: barely "a few hundred" orders placed by AI agents

🔎 Agentic e-commerce makes headlines, the numbers make people smile

Autonomous shopping by AI agents is supposed to revolutionize online commerce. Except that at DoorDash, the American leader in food delivery, the revolution amounts — for now — to a handful of orders.

Speaking during the third quarter 2026 earnings call, CEO Tony Xu stated that DoorDash had received only "several hundred orders" — a few hundred orders in total — coming from third-party AI agents like Muse, Instinct, or Fo. Put that in perspective against the 3 million deliveries per day the company reports in its 2025 annual report. The ratio is so low it borders on a statistical anecdote.

This is the reality check the agentic commerce sector needed. While conferences get carried away about "agent-first shopping," the real data tells a very different story: users continue to order through the DoorDash app, period.

This finding echoes a pattern we have already observed elsewhere in the AI ecosystem: spectacular announcements, disappointing actual adoption figures. As shown in our analysis of Apollo, which still hasn't returned anything to the 493 other companies in the S&P 500, you must systematically separate media noise from measurable traction.

The Essentials

• The actual numbers are tiny: a few hundred orders from third-party AI agents in total at DoorDash (Q3 2026), compared to roughly 3 million daily deliveries through traditional channels.
• But DoorDash is playing the openness card: the company launched a text-to-order agent in September 2026, a CLI agent (dd-cli) in invite-only beta since July 2026, and publishes an official MCP to connect third-party agents to its catalog, cart, and checkout.
• The market is fragmenting: Walmart is opening its channels to personal agents while Amazon blocks them — DoorDash, for its part, is choosing to become a "service layer" rather than a destination.
• Tony Xu mentions advertising in agents as a potential future monetization avenue, but with no established business line at this stage.
• Honest conclusion: agentic commerce infrastructure is being built seriously, but mainstream consumer adoption hasn't followed yet.

Recommended tools

To follow or test agentic commerce yourself, here are the concrete building blocks available today.

If you're new to autonomous agents, our guide to the best autonomous AI agents is a good starting point before connecting anything to a commercial API.

What exactly does DoorDash say? — The sourced figures

DoorDash received "maybe a few hundred orders in total" from third-party agents, according to Tony Xu. That's the exact quote from the Q3 2026 earnings call: "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".

Three sources converge on this figure: DoorDash's official quarterly earnings release, TechCrunch's coverage from October 30, 2026, and the analyses from creati.ai and winzheng.com on the same call.

Which agents are we talking about? Muse, Instinct, and Fo — third-party shopping agents that promise to order your dinner automatically. The result: a few hundred orders. Not a few hundred per day. In total.

Why this figure matters. DoorDash is one of the very first players in commerce to publicly release its agentic usage metrics. Most companies settle for partnership announcements. Here, we have a raw figure, owned up to, by the CEO himself. That's rare and valuable.

And this isn't an isolated case of a gap between rhetoric and reality. Our article on OpenAI offering up to 90% discounts to poach Anthropic and Replit customers shows the same dynamic: a market that gets carried away narratively long before actual usage follows.

What DoorDash Is Building Behind the Scenes — A Deliberate Openness Strategy

DoorDash isn't merely observing that agents aren't ordering: the company is actively building the infrastructure for them to do so.

Three agent products in six months

The timeline is aggressive. In June 2026, DoorDash launches Ask DoorDash, its conversational chatbot. In July 2026, the CLI agent — nicknamed dd-cli — arrives in invite-only beta: a command-line tool that lets an AI agent search for restaurants, compare deals, build a cart, and place real orders. In September 2026, the text-to-order agent makes its debut.

The official MCP: the centerpiece

The most interesting part for developers: DoorDash publishes an official MCP that lets you connect an AI agent to the same catalog, the same cart, and the same checkout as the consumer app. Concretely, your agent doesn't go through a watered-down API — it accesses the real inventory, the real prices, the real payment. The documentation is available through DoorDash's official developer portal.

The strategy is clear, and DoorDash itself has articulated it via Yahoo Finance: to become a service layer that third-party developers and AI systems can plug into, rather than a consumer destination. DoorDash wants to be the payment and logistics rail of agentic commerce, not just an app.

Monetization remains unclear

Tony Xu has mentioned advertising within these third-party agents as a future possibility. But he was explicit: it's not an established business line. In other words, DoorDash is investing in agentic infrastructure without yet knowing how it will pay off. It's a bet on positioning, not on immediate revenue.

Why So Few Orders? — Four Structural Reasons

The short answer: AI agents don't yet solve a real problem for the food delivery consumer.

Reason 1: ordering dinner takes 90 seconds. The DoorDash app remembers your addresses, your favorites, your habits. The journey is already maximally optimized. An agent that "orders on your behalf" saves you... a few taps. The time savings are marginal, while the perceived risk (wrong restaurant, wrong order, payment issue) is real.

Reason 2: trust isn't there yet. Handing a payment over to an autonomous agent means giving it credentials, spending limits, and decision rules. Most consumers aren't ready for that. And they're right to be cautious: agents that act on real systems raise serious security questions, which we cover in our article on AI agent red teaming — a field where weeks of testing can be compressed into a few hours, but where most consumer-facing agents simply aren't tested at all.

Reason 3: third-party agents are niche products. Muse, Instinct, Fo — how many users have even heard of these agents? Adoption of personal agents themselves remains low. You can't have millions of agentic orders if the agents themselves only have a few tens of thousands of users.

Reason 4: the use case is poorly chosen. Food delivery is perhaps the worst playground for agentic commerce: a decision with a strong emotional component (whatever you're craving in the moment), low value per transaction, and high trust friction. Industry observers — including Griddl's analysis of dd-cli and the agentic shift — note that B2B use cases and recurring purchases are far more promising grounds.

Walmart opens, Amazon blocks — The market is fracturing

The competitive landscape deserves as much attention as DoorDash's numbers.

According to information reported by winzheng.com, Amazon is actively blocking personal shopping agents on its platform, while Walmart is opening its door to them. This strategic divergence is the real stake for 2026-2027: every major platform is choosing whether the agent is a customer or a threat.

Amazon's logic is understandable: why let a third-party agent insert a layer between the customer and Amazon, at the risk of losing the customer relationship, the data, and the advertising margin? Walmart's logic is the opposite: capture agentic flows early, even at the cost of giving up the interface, in order to be the default destination for agents.

DoorDash, with its MCP and CLI, is clearly positioning itself in the open camp. It's a risky choice — opening its checkout to third-party agents whose UX and brand it doesn't control — but it's the only choice that makes it possible to capture the agentic market if it takes off.

For developers who want to understand the architectures behind these choices, our article on the 5 AI agent patterns that actually work details the orchestration models that truly dominate in production — and the pattern of "the agent as a customer of a commerce API" is an emerging one.

How to hook into DoorDash as a developer — The express guide

If you want to build an agent that actually orders meals, here's the shortest path.

Step 1: choose your LLM

The DoorDash MCP is compatible with models that support the MCP protocol. For an ordering agent, you don't need the most expensive model: Claude Sonnet 4.6 (Anthropic) or GPT-5.4 (OpenAI) are more than enough for reliable tool-calling. If multi-step reasoning becomes complex (deal comparison, time constraints, budget), switch to Claude Opus 4.7 or GPT-5.5. Our guide on the best LLMs for AI agents compares these models specifically through the agentic lens, including their strengths in tool use.

Step 2: implement the MCP

The official DoorDash documentation, available on the platform's developer portal, describes connecting your agent to the catalog, the cart, and checkout. The crucial point: your agent accesses the same inventory and the same prices as the consumer app. No sandbox API disconnected from the real world.

Step 3: test locally before deploying

A non-negotiable best practice: run your agent locally first. With Ollama, you can self-host open source models like GLM-5 (Reasoning) from Z.AI or Kimi K2.6 from Moonshot AI, and iterate on your agent's decision logic without incurring API costs. Our tutorial on open source AI agents with Ollama locally covers this development phase in detail.

Step 4: set spending guardrails

An agent that can spend real money must have guardrails: a per-order cap, a restaurant whitelist, human validation above a threshold. That's exactly the kind of constraint you can configure in agent frameworks — our guide on configuring OpenClaw: SOUL, AGENTS and Skills shows how to define these behavioral rules in an open source agent.

What DoorDash's Numbers Say About the Market — A Cold Read

Three takeaways, without dramatizing or overselling.

Takeaway 1: infrastructure precedes usage, as always. DoorDash is building rails before there are trains. This is the classic platform pattern: Stripe built the payment API years before one-click checkout became standard. Today's few hundred orders are the entry price for tomorrow's market — if that market exists.

Takeaway 2: the consumer is not the first segment. DoorDash's MCP targets corporate contexts first — ordering from enterprise chatbots, internal assistants. That's where trust friction is lowest: the company assumes the risk, not the employee. The few hundred B2C orders say nothing about the B2B potential.

Takeaway 3: the media hype around agentic commerce is at least two years ahead of reality. The industry's most aggressive projections talk about hundreds of billions of dollars of agentic commerce by 2030. DoorDash's Q3 2026 data — a few hundred orders on a platform handling 3 million daily deliveries — suggests the mainstream adoption curve is starting from virtually zero. This isn't a failure, it's a beginning. But any narrative presenting agentic commerce as a wave that's already here is misleading.

My take, stated plainly: DoorDash is making the right strategic choice by opening up, but the real takeoff will come from agents natively integrated into the assistants people already use — Siri, Gemini, ChatGPT — not from independent third-party agents that nobody has installed.

❌ Common Mistakes

Mistake 1: confusing "available infrastructure" with "real adoption"

Announcements about MCP, CLI, and agentic partnerships create the illusion of a mature market. The solution: always demand real usage figures before citing a player as proof of traction. The DoorDash figure — a few hundred orders — is the honest benchmark to counterbalance the miraculous projections.

Mistake 2: deploying a payment agent without guardrails

Connecting an agent to a real checkout without a spending cap or human validation is a ticking time bomb: one faulty reasoning loop and your agent orders 50 meals. The solution: strict spending caps, logging of every decision, and systematic testing in a local environment before any move to production.

Mistake 3: believing that all agents work everywhere

Amazon blocks personal agents while Walmart accepts them. A shopping agent that works on one platform can be outright blocked on another. The solution: check each platform's policies before promising coverage to your users, and design fallbacks.

❓ Frequently Asked Questions

How many orders has DoorDash received from AI agents?

A few hundred in total, according to Tony Xu during the Q3 2026 earnings call (October 30, 2026, sources: TechCrunch, official DoorDash press release). Compare that to the 3 million daily deliveries announced in the company's 2025 report. The agents in question are third-party agents like Muse, Instinct, and Fo.

How can an AI agent order from DoorDash?

Via DoorDash's official MCP, which connects the agent to the platform's catalog, cart, and checkout, or via dd-cli, a command-line tool in invite-only beta since July 2026. Both channels provide access to the same inventory and the same prices as the consumer app.

Why are the numbers so low?

Four reasons: ordering a meal is already fast via the app, trust in autonomous payments hasn't been established, the third-party agents themselves have a tiny user base, and food delivery is a use case poorly suited to autonomy (an emotional decision, low value per transaction).

Is DoorDash making money from AI agents?

Not yet. Tony Xu mentioned advertising within third-party agents as a future avenue, but clarified that it is not an established line of business. DoorDash is currently investing in the strategic position of being the "service layer" of agentic commerce, without any demonstrated direct revenue.

Is Amazon really blocking shopping agents?

Yes, according to information reported by winzheng.com, Amazon blocks personal shopping agents on its platform, while Walmart welcomes them. DoorDash, with its MCP and CLI, is positioning itself in the open camp. The agentic commerce market is therefore fracturing between closed and open platforms.

✅ Conclusion

The few hundred agentic orders at DoorDash are the industry's most honest figure: agentic commerce is infrastructure under construction, not an established market — and DoorDash is betting on openness to be ready for the day when agents finally place orders en masse. If you want to follow this market with a critical eye, start with our analysis of the best autonomous AI agents to understand what actually works today.