Agency
An AI agent does the work for you
About
An innovation lab working in New York and New Jersey. We teach people to work with AI agents and teams of AIs that check each other, build AI orchestration tools, and are preparing our own AI model specialized in orchestration.
What We Build
Poly A1 — a step-by-step path from a chat to your own team of AIs: an agent that does the work on your computer, and several AIs that check each other’s answers. It amplifies human capability — measured, audited, and governed.
Why
The lab's work is driven by accessibility, safety, and high-stakes decision support. Domains with a high cost of error are our proving grounds — what is hardened there then works for everyone.
Recognition
Project leader Vadym Chernets is recognized by Intel Corporation as an expert and successful practitioner, and released a training handbook with the support of the United Nations.
Architecture
Safety is part of the architecture. The system can improve itself, with 100+ AI self-evolution modules, and the AI keeps its identity through 89 selfhood mechanisms.
What Makes It Different
Patent Portfolio
In July 2025, we filed the first of seven patent applications; by November 17, 2025, one already described more than 100 distinct consensus and divergence methods. The international and national filings cover multi-model consensus of leading cloud and local AI models, accessibility systems, automotive safety, and trust & verification — and were prepared with the lab’s own multi-AI platform.
Three of the seven are international applications under the PCT — PCT/US2025/045268, PCT/US2025/055743, PCT/US2026/012950 — each with an International Search Report and Written Opinion; the rest are U.S. national filings.
How We Work
Proof, Not Promises
The team tests its own claims with pre-registration, placebo controls and falsifiable predictions — and publishes the results regardless of the outcome.
An Unedited Changelog
The evolution changelog is auto-generated and unedited, keeping failures and rejections alongside wins — a system that reports only success cannot be trusted.
Safety as Architecture
Safety is wired into the system's kernel, not written in a policy document: every new capability gain must be matched with a corresponding safety constraint.
- 10+AI models reasoning at once
- 100+consensus & divergence methods
- 7patent applications
- 236,975tests
Platform figures are drawn from an auto-generated FACTS registry (code, config, and the trusted ledger); patent figures — from the filings.
Reviews
I have held several leadership positions within the U.S. Army and the Department of Defense, including the role of Deputy Team Chief at the Defense Threat Reduction Agency (DTRA), where I oversaw international security cooperation.
I had the privilege of working with Dr. Vadym Chernets on a strategic initiative in which he delivered a comprehensive analysis and a program management roadmap. His work integrated deep expertise in artificial intelligence, predictive analytics, and organizational strategy.
The clarity and precision of Dr. Chernets’s analysis were exceptional. His unique ability to translate highly complex AI architectures into practical strategies is not only rare but truly transformative. Dr. Chernets consistently demonstrates global leadership in the field of artificial intelligence.
Josh Mulford
Deputy Team Chief, Defense Threat Reduction Agency (DTRA), U.S. Department of Defense
Dr. Vadym Chernets possesses extraordinary abilities in the fields of data science, artificial intelligence, and strategic systems analysis. I had the opportunity to collaborate with Dr. Chernets and observe the exceptional quality of his work and its impact on organizational strategy and innovation.
I spent 20 years on Wall Street, and from my practical experience I can say that what sets Dr. Chernets apart is his rare ability to integrate complex academic and technical concepts into clear, practical strategies. His work has brought measurable value to our organization and transformed our approach to decision-making, resource planning, and innovation.
Boaz Rahav
Chief Financial Officer and Co-Founder of Hypersonic Force, Member of the School of Management Executive Council at New York Institute of Technology
More reviews
Two pillars
Don't learn to code. Learn to direct.
Orchestration
Several AIs work as one team
Pillar 1 · Agency
A chat answers. An agent does
It works in your files, step by step — and waits for your OK.
Pillar 2 · Orchestration
Several AIs. One checked answer
AIs from different companies check each other’s work: one builds the base, the others make it stronger.
Verified answerSee who added what
How pros work: Claude Code builds the base, ChatGPT gives a second opinion, Meta AI a third on hard tasks. Plus 10+ more AIs when needed.
Pillar 2 · Orchestration
One goal. Different roles
Not one AI for everything, but a team: each one does what it does best.
ResearcherFinds the facts
Searches and brings back sources.
DoerDoes the work
An agent on your computer.
Your goalOne task —
many minds
many minds
CriticTries to break it
Looks for weak spots.
CheckerChecks the result
An AI from another company.
The point is not more AIs. The point is AIs that work well together. The arrangement is the craft.
Pillar 2 · Orchestration
What a team of AIs gives you
One agent is useful. A team is powerful.
- 01
Accuracy
They catch each other’s mistakes —
DraftCold criticAnother AIChecked
- 02
Breadth
Many angles. Nothing missed.
ClaudeChatGPTGeminiAll the ideas
- 03
Depth
Research, not a quick answer.
Deep searchFollow-up questionsSources
- 04
Economy
Full power. Smallest bill.
Routine → a fast AIHard → the strongestSubscriptions you already have
What you can hand over
The agent does itSeveral AIs check and add
- Inbox sorted, replies drafted
- Meeting → summary and tasks
- Client offer from your prices
- Receipts in one table
- Folders tidied and renamed
- Second opinion on a contract
- Where AIs agree and argue
- Ideas from many angles
- Deep research on a question
Two worlds are merging
Code was for programmers. Agents are for everyone
Agents used to be for programmers, and chat for everyone else. In 2026 the wall came down.
- 2025
Two worlds
Programmers worked with agents. Everyone else asked ChatGPT and copied the answers by hand.
- 2026
The line blurs
Agents come to ordinary computers. They open files and the browser and call other AIs. You talk to them the way you talk to ChatGPT.
- 2027
An everyday skill
Directing agents becomes a normal skill, like email or spreadsheets. Most people haven’t started yet. You can be early.
The path
From zero to your own team of AIs
One step at a time. You start with the chat you already use.
Your usual chat sets up the agent
ChatGPT, Gemini or Meta AI — whichever you already have. It helps you install Claude Code on your computer. Programmers work with this agent. You don't need code: you talk to it in your own words.
The agent takes your hand
From here the agent leads you itself: gets to know you, teaches, trains. We give it the lessons and steps as ready text — you just copy it in.
Agency
You hand over real work: files, documents, search. You learn to set a goal and accept the result.
Orchestration in one window
The agent opens other AIs in your browser, asks them and brings back the answers. One window, many minds.
A second voice and more
A second agent through a subscription you already have. Two agents check each other. More agents join when needed. Your subscriptions become tools, not a line on your statement you forgot about.
Your phone is the remote
The computer at home works, and you run it from your phone. When you’re back, everything is done.
The path keeps growing. AI changes every week: we test what’s new and add the steps that work.