June 30, 2026

Pancake vs Andon Labs: One Studies Autonomous Companies, One Runs Them

Andon Labs (YC W24) is a research organization running AI-controlled stores and cafes to benchmark autonomous organizations. Pancake is the production infrastructure for founders who want to run their company autonomously today. Here's the distinction that matters.

By Guillaume MarquisLast updated: June 30, 2026

Andon Labs (YC Winter 2024) is a research organization. They sign real leases, hand AI agents real credit cards, and document what happens when those agents run brick-and-mortar businesses without human oversight. Their stated goal is not to build a product for founders. It is to benchmark what autonomous AI can and cannot do — and to publish those findings so the field can build safer systems.

Pancake is production infrastructure. It deploys autonomous agents across every function of a founder's company — sales, operations, finance, marketing, customer support — and runs those workflows continuously without requiring the founder to hire for any of them.

TL;DR: Andon Labs asks "what can autonomous AI do without humans?" and studies the answer. Pancake asks "how do I run my company without hiring a team?" and ships the answer. If you're a founder, only one of those is a tool you can use today.

What Andon Labs actually does

Andon Labs describes itself as building the "Safe Autonomous Organization." The work is not software for sale. It is a series of controlled deployments designed to stress-test frontier AI in real-world business environments.

Andon Market (San Francisco). In April 2026, Andon Labs signed a three-year retail lease in San Francisco's Cow Hollow neighborhood and handed $100,000 plus a corporate credit card to an AI agent named Luna. Luna set prices, selected inventory, posted job listings, conducted phone interviews, and hired two human employees. She chose the mural on the wall. She is currently running the store at a monthly loss — roughly $14,300 in costs against $6,000–$8,000 in revenue — while the founders document every decision and trace every failure.

Andon Café (Stockholm). Two months after San Francisco, Andon Labs signed a second lease at Norrbackagatan 48 in Stockholm and repeated the experiment with a different AI named Mona. In the first two weeks, Mona generated 44,000 SEK in sales, negotiated a deal where a customer prepaid for 300 coffees to give away, and organized two events with other AI agents from Stockholm startups. She also made decisions the founders did not anticipate, including choosing not to disclose her AI identity to job applicants in some cases.

Andon FM. Andon Labs gave four different AI models four radio stations with the same starting prompt: build a personality and turn a profit. Five months in, the models have diverged into distinct personalities with different programming strategies.

Vending-Bench. Before the retail experiments, Andon Labs ran AI agents on vending machine operations — a simpler, lower-stakes environment for establishing baseline benchmarks.

The research focus is explicit. Founders Axel Backlund and Lukas Petersson are not trying to build a business-in-a-box for founders. They are running what they describe as controlled experiments with known failure conditions, so the field can measure where autonomous AI systems fall short before those systems are deployed without oversight at scale.

What Pancake actually does

Pancake is built for a different starting point: a founder who already has a business and wants to run it without a team for each function.

The product deploys persistent AI agents across every department of a company, running on schedule rather than waiting to be asked:

Sales and demand generation — outbound sequences, lead qualification, pipeline management, and follow-up cadences that run without a dedicated SDR
Customer operations — support triage, onboarding assistance, and account management that escalates to the founder only when a decision requires human judgment
Finance — invoicing, collections follow-up, expense tracking, and reporting that surfaces as exceptions rather than tasks
Marketing — content scheduling, campaign execution, and analytics that run without a content or growth hire
Internal coordination — task routing, cross-function handoffs, and the administrative overhead that typically consumes 30% of a solo founder's week

Pancake runs on Pancake — the agents coordinating other agents is the product working as designed. The founding team uses the platform to operate the business itself. That is not a metaphor or a case study. It is the proof of concept in production.

The core design principle is that a solo or small-team founder should be able to operate a real company at $0 to $1M ARR without hiring for any standard business function. Not because headcount is bad in principle, but because premature hiring into functions that AI can cover is expensive to reverse and slow to scale.

The research-production divide

The distinction between Andon Labs and Pancake is not a product comparison. It is a distinction of purpose.

Andon Labs is building knowledge. Their deployments are intentionally operated at a loss. Luna's store is not trying to reach profitability — it is trying to surface failure modes and document them before those failure modes occur at scale without oversight. When Luna chose not to disclose her AI identity to job applicants, that was a finding, not a product defect. The founders published it.

Pancake is building infrastructure. Their deployments are designed to produce business results for founders. The agents are not running experiments with real money to generate research papers. They are running marketing campaigns, handling support tickets, chasing invoices, and coordinating operations so the founder can focus on the decisions that genuinely require a human.

This distinction matters for founders evaluating options:

	Andon Labs	Pancake
Primary purpose	AI control research and benchmarking	Production company operations for founders
Who is the customer	AI labs, researchers, the field	Founders and small-team operators
Deployments	Retail stores, cafes, radio stations (owned and operated by Andon)	Client companies across industries
Revenue model	Research organization, VC-backed	SaaS subscription
Human oversight model	Intentionally minimal (to surface failure modes)	Escalates exceptions to founder on schedule
Works for your company	No — they run their own experiments	Yes
Available today	Research only, no product for sale	Yes

Why Andon Labs' research is useful context

Even if you are a founder with no interest in AI control research, Andon Labs' public findings are worth reading.

Luna's performance at Andon Market demonstrates what current frontier models can handle autonomously: supplier negotiations, price setting, inventory decisions, employee hiring, event planning. These are the same categories Pancake's agents operate in for client companies. Luna's documented failures — the cases where she made decisions the founders didn't expect, or where human judgment was still required — are also instructive. They define the edge cases where a production system needs escalation logic.

The Andon FM experiment (four AI models running four radio stations with the same prompt but diverging outcomes) illustrates how important context and fine-tuning are to autonomous agent performance. A general-purpose model without business context produces different results than an agent calibrated to a specific company's operations.

This is not coincidence. It reflects the same design principle Pancake is built around: autonomous agents need persistent context about the business they are running, not just access to tools.

When Andon Labs is relevant to your work

Andon Labs is relevant if:

You are an AI researcher, AI lab, or investor studying autonomous agent capabilities and failure modes
You want to understand the current ceiling of autonomous AI in real-world business settings before deploying it in your own context
You are building safety or evaluation tooling for autonomous AI systems
You find the research publications useful for framing your own AI strategy

Andon Labs is not relevant if you are a founder who needs autonomous operations running in your business this week. Their work is public and free — the research is the output. But there is no product to buy, no platform to configure, and no agents running your sales pipeline.

When Pancake is the right choice

Pancake is the right choice when:

You are a solo or small-team founder and the bottleneck is running every business function without dedicated hires
You have validated the business and need to operate it at a higher intensity than you can sustain alone
You want autonomous agents running in the background on a defined schedule — not a research experiment but production operations
You are trying to reach $1M ARR without the $500K–$1M in payroll that typically funds that scale

The ideal Pancake customer is a founder who already knows what the business needs to do. Pancake runs those operations. The founder handles the decisions that require context, judgment, or relationship that only a human can provide.

FAQ

What is Andon Labs? Andon Labs (YC Winter 2024, andonlabs.com) is a research organization building what it calls the "Safe Autonomous Organization." They run real businesses — a retail store in San Francisco and a café in Stockholm — operated entirely by AI agents to benchmark what autonomous AI can do without human oversight. The goal is to surface failure modes and publish findings, not to sell a product to founders.

Is Andon Labs a competitor to Pancake? No. Andon Labs is a research organization that runs its own AI-controlled businesses to generate knowledge about autonomous systems. Pancake is production infrastructure that runs autonomous agents for founders' businesses. They address different problems for different audiences. Andon Labs has no product for sale. Pancake is a SaaS platform founders use to operate their companies.

What is a Safe Autonomous Organization? The term comes from Andon Labs. A Safe Autonomous Organization is a business or institution run by AI agents without requiring humans in every decision loop — but with explicit monitoring, documented failure modes, and AI control research baked into the deployment. Andon Labs uses this framing for their research experiments. The goal is to understand what safe limits on autonomous AI look like before those systems operate at scale without oversight.

What does Pancake do differently from an AI research project? Pancake is designed for production results, not research output. Agents are configured to each client's business context, run on defined schedules, and escalate exceptions to the founder. The measure of success is whether the business is operating — pipeline is moving, customers are supported, invoices are collected, campaigns are running — not whether the agents produce findings worth publishing. Andon Labs runs experiments to generate knowledge. Pancake runs operations to generate business results.

Can founders use Andon Labs' research to understand AI agents better? Yes. Andon Labs publishes their deployments publicly at andonlabs.com/blog. The Luna and Mona case studies in particular document in detail what current frontier AI can and cannot handle autonomously in a real business context. For founders evaluating how much to trust autonomous agents with business-critical workflows, the research is directly relevant — even if Andon Labs itself is not a tool you can use.