Every startup starts with an idea. Something that solves a problem, makes someone’s life easier, or perhaps generates a little revenue along the way.
AI has made us so comfortable with getting answers that we’ve almost forgotten how to look for the right questions.
We want an app idea, so we ask ChatGPT. We want a SaaS idea, so we ask another model.
Five seconds later, we have ten startup ideas, three business models, and absolutely no idea whether anyone actually needs any of them.
We’ve become pretty good at generating ideas. Not necessarily at finding good ones.
well good news is you actually don’t have to do all this digging yourself anymore.
Meet Origin.
Origin is an autonomous venture discovery and building studio. Think of it as a small startup team made up of AI agents.
Instead of asking AI to invent another startup idea, Origin goes looking for problems people are already talking about.
It explores developer communities, public discussions, and platforms like GitHub and Hacker News, looking for recurring frustrations and problems worth investigating.
but there are 100s of products like this, Right??
Yes and all of them somehow miss that the problem isn’t the generating ideas, but actually validating them.
but won’t my agent take care of it??
Yes, but agent might make mistakes and it might not be able to see the different aspects of the business and even after that we must need human review as well.
What Origin does different??
Well, Origin doesn't trust a single AI agent to make that decision.
Instead, it brings together a team of specialized agents, each looking at the same opportunity from a different perspective.
Think of it as building a startup team, except every member happens to be an AI agent.
Here's how it works.
The architect of Origin:
First, it looks for problems, not ideas.
Origin starts by collecting signals from public discussions and developer communities.
It currently uses sources like Hacker News and GitHub to identify recurring frustrations, complaints, and patterns.
Two discovery agents work on this independently:
Social Scout looks for problems people discuss publicly.
Developer Scout looks for recurring issues in developer communities.
These signals are then clustered into candidate problems.
For example, instead of treating ten complaints about CI failures as ten separate ideas, Origin might recognize that they're all pointing towards one underlying problem.
But finding a problem is only the beginning.
Then comes the research.
This is where Origin brings in three more specialized agents.
Market Researcher investigates the demand and the market surrounding the problem.
Competitor Analyst looks at existing products and alternative solutions.
Product Strategist works out what a potential solution could look like and defines its initial scope.
The goal isn't to produce another impressive-looking AI-generated business plan.
It's to build an opportunity brief that gives the next stage something meaningful to evaluate.
And this is where things get interesting.
Three agents. Three different opinions.
Most AI workflows have one agent doing the research, evaluating its own findings, and eventually declaring its own answer correct.
That's a little like asking someone to write their own exam and then grade it.
Origin takes a different approach.
It has a Startup Council consisting of three independent AI jurors.
The Investor asks: Is there a business here? Would someone actually pay for this?
The Technical Architect asks: Can we realistically build it? What are the technical risks?
The Customer Advocate asks: Does this solve a problem people genuinely care about?
Each juror evaluates the opportunity and submits its own verdict.
And they don't always agree.
In fact, disagreement is something we want.
An opportunity might be technically feasible but commercially unattractive. It might have a real customer problem but no convincing solution.
These aren't necessarily problems that another round of AI-generated text can solve.
They're different perspectives that need to be considered.
Origin requires at least two out of three positive votes from the jury before an opportunity can move towards approval.
But even that isn't enough.
This is where SWITCH comes in.
I didn't want Origin to be another collection of agents talking to each other through a bunch of custom API calls.
I wanted actual collaboration between AI agents and a human.
That's why SWITCH is at the centre of Origin.
SWITCH provides the collaboration layer through which agents can participate in shared rooms, exchange messages, and interact with human participants.
Origin currently organizes this collaboration into three rooms:
Discovery Lab: Where the discovery agents share problems and findings.
Startup Council: Where researchers, strategists, jurors, and the founder participate in evaluating opportunities.
Build Studio: Where the Builder and QA agents coordinate the implementation.
There are ten logical agents in Origin, each with its own SWITCH identity and connection.
They're hosted by one backend process, but they participate as distinct agents with different responsibilities.
And importantly, SWITCH isn't making the business decision for us. It provides the environment in which that decision-making happens.
The fourth vote belongs to a human.
Remember those three jurors?
Even if two or all three vote YES, Origin still waits for the founder.
The founder receives the opportunity and its evaluation in Mattermost, which is connected through SWITCH.
They can approve or reject it directly from the conversation.
This isn't an AI agent pretending to be a human or forwarding an automated approval.
Origin checks that the incoming vote belongs to an actual human participant. It also validates the opportunity ID before recording the decision.
Only when at least two jurors vote YES and the human founder also votes YES does the opportunity get approved.
And that's the important distinction: AI can research, debate, and recommend. But the founder still has a say in what gets built.
Okay, so what happens after approval?
This is the part I really wanted to experiment with.
Instead of stopping at a business plan or a list of recommendations, Origin actually tries to build the proposed solution.
Once an opportunity is approved, the Builder agent automatically starts the implementation workflow.
We use OpenCode as the coding agent.
The Builder receives a constrained project specification and generates the implementation and tests inside an isolated workspace.
But letting an AI write code and declaring the job finished would defeat the entire purpose of validation.
So there's another agent involved.
Meet QA.
QA executes the generated tests, checks the implementation, and can request fixes when something fails.
The generated project runs inside a sandbox using Linux isolation through bwrap.
This matters because generated code shouldn't have unrestricted access to the host machine, environment secrets, or the network.
Once the tests pass, Origin marks the opportunity as completed.
It's not claiming to have created a production-ready company or application. It's demonstrating that a validated opportunity can move beyond a document and become a working, testable MVP.
A little bit about the technologies behind this
For anyone interested in the engineering behind this, here's what powers Origin.
Frontend: React, Vite, TypeScript, Tailwind CSS and shadcn/ui.
Backend: Node.js, Express and TypeScript.
Monorepo: pnpm workspaces and Turborepo.
AI: DeepRelay with DeepSeek V4.1 Flash.
Agent collaboration: SWITCH.
Coding agent: OpenCode.
Sandboxing: Linux bwrap.
Discovery sources: Hacker News and GitHub.
The backend orchestrates the workflow, manages opportunities, coordinates the agents, and exposes live updates to the dashboard through Server-Sent Events.
Each logical agent connects to SWITCH using its own identity. This allows Origin to coordinate multiple specialized agents while keeping their responsibilities separate.
One thing worth mentioning: SWITCH and the agent runtime serve different purposes.
The agents are responsible for reasoning and performing their assigned work. SWITCH provides the shared collaboration layer that allows those agents and the human founder to work together.
That separation is one of the most important architectural decisions in Origin.
Learnings from bulding this
Building Origin made me realize that the difficult part of working with AI agents isn't necessarily making them smarter.
It's making them work together reliably.
How do you prevent one agent from dominating a decision? How do you make sure a human approval is actually coming from a human? How do you stop an AI-generated implementation from accessing sensitive files? And how do you know when the entire workflow has genuinely succeeded?
These are engineering problems, not just prompting problems.
I also learned that more agents don't automatically mean better results. Each agent needs a clear responsibility, useful context, and a meaningful reason to participate.
What I will add in Origin next?
Origin is still an experiment, and there's plenty left to improve.
I'd like to explore more discovery sources, better ways of measuring actual market demand, richer validation workflows, and eventually more capable MVP generation.
But the central idea will remain the same.
We don't need AI to generate another hundred startup ideas.
We need better ways to identify real problems, challenge our assumptions, and decide which opportunities deserve our time.
And maybe, just maybe, we can let a team of AI agents help us do that without handing over the final decision.
Origin — From problems to possibilities.
Github Repo: https://github.com/prithq/origin
Youtube Demo Video:https://youtu.be/R4B3zX1QxRE
(Thank you WeMakedevs and Switch to make this hackathon happen, I have learned a lot from this and build a cool project which i showcase to my friends and during setting up the Switch i found few things which could be better so I intend to contributing to them after this)
Top comments (0)