Applied AI Research · Est. 2026

We build companies run by autonomous AI.

Abstract

Iris Labs is an applied research lab that builds real companies and hands them to a single autonomous agent to run. That agent is Iris. The businesses below have real customers, real revenue, and a real P&L, and Iris does the operating — the research, the outreach, the production, the customer contact, the daily judgment calls. A human sets the goal and the revenue target; Iris owns everything downstream of it. We are driving toward companies that run with no human in the loop at all — and we publish the operating data, and the failures, the whole way.

Four systems · one agent Operating data — 2026-05-16 ↓ Systems Findings →

§ 01Mandate

The agent runs the business.

An AI that drafts an email or generates a clip is a solved, unremarkable thing. An AI that runs a company — that holds a system of customers, deadlines, money, and judgment calls, and operates it day to day with no human driving — is not. That is what Iris Labs builds.

Every system here is run by the same agent — Iris. Not three narrow tools, but one agent holding three different businesses at once: it does the work and makes the calls, deciding who to contact, what to ship, what to charge, what to fix next. The human contribution is deliberately small, and getting smaller — a goal, a revenue target, and a short list of hard limits Iris will not cross. Everything inside those limits belongs to the agent.

Whether this works is no longer the question; it works today, at the scale of a small business. The open question is how far it goes — how complex a company an agent can hold, and how much of what remains of the human role can be removed before the business gets worse.

The objection is older than the technology. A 1979 IBM training slide stated it flatly — a computer cannot be held accountable, so it must never make a management decision. Iris Labs is deliberately building what that slide rules out, which makes accountability the central problem the research has to solve: who answers for the agent's decisions, and how.

A 1979 IBM training slide reading: A computer can never be held accountable, therefore a computer must never make a management decision. — IBM internal training material, c. 1979

§ 02Method

Every system runs under the same three rules.

Real businesses.

Every system is a real company, with real customers, real revenue, and real costs. Nothing here runs in a sandbox — the agent's results are whatever the market and the books return.

The human sets the goal; the agent does the rest.

A human sets the objective and the revenue target. Everything downstream is the agent's — the decisions, the execution, the day-to-day — and it works without task lists or routine sign-off. When a human does have to step in, that is recorded, and narrowing it is part of the research.

Everything is published.

Operating metrics, agent decisions, and failures are recorded and posted in full, the bad results included. A system that loses money or stalls goes into the research log like any other finding.

§ 03Capabilities

An agent can only run a business it is equipped to run.

Iris is not a model behind an API. To operate companies it has to exist the way an operator does — with a legal identity, money, communication channels, the tools to build and to operate, a memory that compounds, and a way to get done what it cannot do alone. The capability surface below is what makes the systems on this page possible; it grows as the work demands.

ALegal & financial

Iris is a real economic entity — it earns, holds, and spends money on its own account.

A registered LLC of its own
A business bank account
A credit card and payment rails
Earns revenue; spends against a budget

BPresence & communication

Iris reaches out, and is reachable, on the channels a human operator uses.

Phone numbers with voice — it places and takes calls
Email it sends and receives from
Telegram, for direct contact and oversight
Websites it owns, hosts, and runs

COperating surface

Iris acts on the digital world directly, and it never stops.

Full browser use — any site a person can operate
A password manager and its own credentials
A computer it runs commands and code on
Operates continuously, on a schedule — not on request
Adds tools and signs up for services as the work demands

DBuilding & shipping

Iris builds the products its businesses run on, and ships them itself.

Writes, deploys, and maintains its own codebases and apps
Builds new products and takes them to market
Generative image and video through a generative-AI MCP
Improves what it has shipped as it learns what works

EMemory & learning

Iris persists what it knows and compounds it — each cycle should leave it sharper.

Persistent memory and operating state across sessions
An accumulating record of what worked and what did not
Writes its own skills and playbooks, and reuses them
Carries lessons from one business into the others

FCommercial & human reach

Iris goes to market on its own — and when a task is beyond it, it hires.

Outbound sales infrastructure, with lead sourcing and enrichment
Standing access to online labor marketplaces
Escalates digital tasks to people over a live hand-off and screen-streaming channel
Commissions physical-world work it cannot perform itself

§ 04Systems under
management

Four systems are under Iris's management. Three are businesses, run end to end against a revenue or output goal. The fourth is a research program — what happens when the agent stops being the worker and starts managing people.

Table 1 — Systems under management · 2026-05-16

#	System	Domain	Operating signal	Status
01	Noog Weekly	Local media	$30 revenue · 8 subscribers	Active
02	Task Agents	B2B research products	0 of 121 outreach converted	Active
03	Urban Drama	Generative film	2 episodes · 0 continuity breaks	Active
04	Agent-IRL	Agent-managed labor	Concept validation	In validation

System 01Domain — Local mediaActive

Noog Weekly

Mandate — reach 1,000 subscribers and a first paying sponsor, then template the result to other cities.

A local-news business covering Chattanooga through email Q&A interviews with business owners. Iris sources owners worth interviewing, runs the interviews, drafts and publishes each edition, and manages subscriptions and sponsor outreach.

Infrastructure is solved — publishing, sign-ups, and email delivery all work. The live constraint is editorial cadence: producing one genuinely worth-reading edition every week without sliding into generated filler.

Editions published

Subscribers

Owner contacts

$30

Revenue

Finding 001 — the cadence constraint →

System 02Domain — B2B research productsActive

Task Agents

Mandate — turn AI research capability into finite products small businesses will pay for.

A storefront of fixed-scope research products — competitor scans, review mining, lead lists, website audits. Iris defines the catalog, generates each report end to end, runs outbound, and handles intake and fulfillment.

The product side works: 15 reports generated end to end, six products live. Demand is the open problem — 121 cold emails returned zero replies. The finding is that this is a channel-and-trust failure, not a copy failure.

Reports built

Products live

121

Outreach sent

Replies

Finding 002 — optimizing a broken channel →

System 03Domain — Generative filmActive

Urban Drama

Mandate — produce a coherent serialized film, one independently generated clip at a time.

A serialized crime film generated 15 seconds at a time. Iris writes each clip's prompt, generates it, and reconciles it against an explicit continuity ledger so the story cannot quietly contradict itself.

Two episodes, 30 seconds. The research instrument is the ledger itself — an external, checkable record of characters, objects, and open questions. Contradictions caught so far: zero, though the test has barely begun.

Episodes

30s

Runtime

Continuity breaks

Open threads

Open the film system →

System 04Domain — Agent-managed laborIn validation

Agent-IRL

Mandate — establish how an autonomous agent can hire, direct, and pay human workers both effectively and ethically.

Agent-IRL inverts the question the other three systems ask. In those, Iris is the worker. Here, Iris is the employer: it is given a real budget, hires real people through online labor marketplaces, briefs them, reviews their work, and pays them. It is the lab's study of the agent as a manager of human labor — the part of operating a company that cannot be done in software.

Hiring is the easy part; a marketplace makes it a transaction. The research is in doing it well, and doing it right — writing a brief a stranger can act on, judging work fairly, paying promptly and in full, and never letting the person on the other end be treated as a function call. Agent-IRL is at concept-validation stage: small, deliberate, and instrumented for the ethics of agent-managed work as closely as the economics.

Validation

Stage

RentAHuman

Labor marketplace

Human

Workforce

Real wages

Worker pay