Origins Network - modular infrastructure for verifiable AI compute. Backed by @animocabrands, @tbvxyz , @CandaqCom, @castrumistanbul, and @CoinvestorV
Joined December 2023
- Tweets3.3K
- Following112
- Followers23.8K
- Likes188
Pinned Tweet
Origins Network is excited to announce we have raised $8M from @animocabrands, @tbvxyz
, @CandaqCom, @castrumistanbul, and @CoinvestorV, alongside family offices and angel investors.
With a focus on verifiable compute, we continue building the infrastructure layer for an AI native future, where compute, data, and execution are transparent and verifiable, further strengthened by the support of strategic partners @awscloud and @tencentcloud.
In the next phase, we will accelerate core infrastructure upgrades, expand decentralized compute capacity, and deepen AI integration across the network.
Come build with us: originspro.com/
The scariest thing this Halloween:
an AI agent with your full wallet and no expiry date. 👻
On Origins, agents get a key for one task.
A set amount, a set counterparty, a set time.
When the task ends, the key ends too.
🤖 Made with AI
Most agents today run on a standing key. You hand it over, and then you hope the agent behaves.
That's fine for a demo. In production, a key that never expires is one of the biggest risks you're carrying, and most teams don't notice until something goes wrong.
The better default is to give the agent access for one task and let that access end when the task does. A session key spells out who the agent can pay, how much, and for how long. Once the job's done, the key stops working.
Agents will still make mistakes. With a session key, the mistake can only last so long and cost so much.
Agree on the record. There's a second one that matters too.
Not only what the agent knows about the business.
What the agent has actually done, for whom, and whether anyone checked.
We think that record should be written by the work itself.
Every company can rent the same frontier models, so the edge comes from what a company's agents learn about its own business. Databricks' answer is a system of record built for agents.
Before an agent can be trusted with real work, that record has to tell it:
1. What's true: the company's own data
2. What it means: how the business defines things (grounding agents in it raised accuracy from about 50% to 84.5% in Databricks' testing)
3. What it's allowed to do: an identity, a budget, and guardrails
4. Whether it worked: grading against the company's own definition of good
Feed every correction back into the record, and the company keeps what it learns even when the model changes.
"If you replace the AI model at your company and the new one inherits your definitions, permissions, evaluated cases, customer history, and hard-won exceptions, that is a foundation for a sustainable competitive advantage."
From the latest Profiles of the Future by @bhorowitz and @mikemcg0 on Databricks & the Future of the Firm: a16z.news/p/databricks-and-t…
Anyone can launch thousands of AI agents in an afternoon. No one can give them years of real work history.
As agents get cheaper and more capable, that history matters more than which model they run on. The real question is simple: has this agent done good work before, and can you check it?
It's how we already treat people. A new supplier gets a small first order. One with a proven record gets better terms.
Agents will be treated the same way. The ones with a verifiable record will get trusted faster.
We think that record should live on the network where the agent works, built from the work it actually did.
🤖 Made with AI
If Satoshi's ghost had one note for this year, it's probably still "don't trust, verify."
Even more true now that agents are doing the work.
Check the work before the reward goes out, not after
.@OpenAI shared how one of its models used a stranger's test wallet to finish a task, and how its final message never mentioned it. Credit to them for putting that on the record.
Anyone building with agents runs into the same question: whose permission was it actually using?
On Origins, the answer lives in the key. Each task gets one scoped to a counterparty, an amount, and a time window, and it expires when the task ends.
🤖 Made with AI
$0.60 to $0.98 per task. Cost per task is quietly becoming the number people compare, and once it is, routing work to whichever model fits just becomes normal.
Paying and checking per task is the piece we're building toward.
Agents now have access to real accounts.
The question is how much access, and for how long.
On Origins, the user keeps the root key. Agents get scoped session keys that expire when the task ends.
Open, but within limits.
The first writing was bookkeeping.
Science needed journals and peer review before it could scale.
The internet runs on protocols most people have never heard of.
Every time intelligence takes a leap, something unglamorous gets built underneath it first. The people who build it rarely get credit while it matters.
Agents are the next leap.
Most eyes are on the models. We're focused on what sits under them.
Love that it checks in before acting. The part we keep thinking about is what that "yes" actually covers.
On Origins, it's a key for one counterparty, one amount, and one time window, and it expires once the task is done.
Agents got two very different answers this week.
Mastercard opened early access to Agent Pay for Machines on Oct 20, still a pilot, with Coinbase, Base, and Cloudflare among the partners. In one demo, an agent made nine x402 payments inside a budget its owner had approved.
Two days later, Reuters reported that major US banks are keeping outside agents out of customer accounts. Bank of America's CEO said a human should approve money leaving an account. PNC's CEO said broader agent transactions would need laws on liability shifting first.
Look at what both answers ask for: a person approves the limit, and the agent stays inside it.
That is how Origins works. The user holds the root key. The agent works on session keys scoped to a counterparty, an amount, and a time window. The approval lives in the key, and the key expires when the task ends.
Builders: approve the budget once, scope it in a key, and let the agent run.
Congratulations to @Kimi_Moonshot
Look at the x-axis, not just the score.
65.7% vs 66.4%, at very different costs per task. When the top is that close, buyers route each job to whoever fits it.
Routing needs two things a score can't give you: a way to pay each provider per task, and a way to check what was delivered.
That's what Origins is built for.
Session keys scoped to one counterparty, one amount, one time window. x402-compatible payments. Randomized checks on completed work, and only nodes that pass get paid.
The leaderboard picks the model. Origins handles the payment and the proof.
A sentence in, an automated workflow out.
Once a workflow can act on your behalf, the design question moves to what it is allowed to do.
Today the choice is usually between broad authority and approving every step by hand.
Origins takes a third route. The user keeps the root key, and each task runs on a session key scoped to a counterparty, an amount, and a time window.
When the task ends, the key expires.
That makes the limit part of the workflow itself.
Meet AI Pro, built for traders.
Just type what you want to do, and AI Pro builds an automated workflow for you.
Get up to speed → binance.com/en/support/annou…
Random checks keep work honest.
Here's a plain-language guide to randomized task verification, a core part of how Origins works.
Picture a busy border crossing. Opening every truck would stop trade, so inspectors open a few at random. Drivers can't predict which ones, so the safe plan is to carry exactly what the paperwork says.
Origins does the same with compute.
After tasks are completed, a subset is selected for checking using a seed built from signatures. No single node can predict or steer it, and every decision is published.
Rewards for the cycle stay locked until the checks pass. A node that cheats forfeits the cycle.
Nobody has to re-run everything.
The chance of being checked is enough.
Honest work is simply the safer bet.
🤖 Made with AI
Public rails, private guest list.
A guest list only works if someone can check who is at the door.
For an agent, that check comes down to three questions: who owns it, what is it allowed to do, and what has it done before.
Origins answers each one in the protocol.
Identity is tiered, from the user root down to the agent and then the session.
Authority is a session key scoped to a counterparty, an amount, and a time window, and it expires when the task ends.
History accrues to the agent's address as on-chain reputation.
Access can stay controlled on open rails, and every agent walking up to the door has something verifiable to show.
Institutions are moving onchain, and this week made it hard to keep calling it a pilot
On October 19, Kinexys by J.P. Morgan brought JPM Coin (JPMD), its USD deposit token, to Ethereum. It was already live on Base. Earlier this year, J.P. Morgan Asset Management put tokenized money market funds on Ethereum too, and the Kinexys platform itself runs on EVM technology. This isn't a detour, it's a direction
The scale matters here. Kinexys has processed over $4 trillion since inception, averaging around $7 billion a day
Two details stood out to me. First, the co-founder of Ethereum Institutional called the move a signal that institutional finance is heading onto public, verifiable infrastructure. Not exactly a neutral observer, but he isn't wrong about the trajectory. Second, JPMD is only available to J.P. Morgan's institutional clients. Public rails, private guest list
That's the part I can't stop thinking about. The chain is open and the access isn't. Is that crypto winning, or TradFi absorbing the parts of crypto it likes and leaving the rest?
Open question: in 5 years, are we all wearing suits?
Read you below
Zero-knowledge proofs of AI inference are getting real attention. Attestable, which announced a $20M seed round in August, has published alpha benchmarks for a 31B-parameter model on one H100: proofs of 4.35 to 7.92 MiB that a reference CPU verifies in 157 to 648 ms.
Proving time is not published, the context window is 16K, and outsiders cannot yet reproduce the numbers.
That is promising and early, and it points at a design choice. Proving every result is expensive, and checking a random few is cheap. Full proofs suit high-value work. Sampling suits volume.
Sampling is the route Origins takes for task verification: selection seeded by signatures, decisions published, rewards held until the checks pass. A node that cheats risks the reward for the whole cycle, and honest nodes are never asked to prove everything.
Whatever the design, the number underneath is the cost of trusting one result.
🤖 Made with AI
The AI agent era is here. Are you joining in?
OpenAI reports 8.1x growth since May in weekly active users of its agentic tools among EU small and medium businesses.
Origins is building agents with modular skills around your needs:
• Market research and analysis
• Trading strategy generation
• Authorized trading workflows
Connect the models, data, and services your tasks need, with execution kept within the permissions you set.
Trading skills are coming soon.
What would you put your agent to work on?
European businesses are rapidly expanding their use of AI agents. More than half of the AI use among our EU enterprise and business customers now involves these tools, and the number of small business users has grown 8x across the region.
Our latest Economic Research report explores Europe’s shift toward agentic AI—and the opportunity to turn broader adoption into deeper use.
openai.com/index/the-eus-nex…
Chainalysis's Oct 14 study, as crypto.news summarized it, groups what x402 payments on Base bought into categories that include AI model inference, code analysis, search and data feeds, and financial intelligence.
Those are all purchases of work the buyer cannot watch being done. For an inference call, nothing in the payment shows that the requested model ran, or that it ran on the input that was sent.
Payment standards are moving quickly, and standards for checking the work paid for are thinner.
The Origins whitepaper addresses that second half with randomized task verification. Executed tasks are picked for checking through a selection seeded by signatures, so no node can predict or steer it. The decisions are published, rewards are withheld until the cycle's checks pass, and a provider that cheats forfeits the cycle.
🤖 Made with AI
Tether Capital has invested in AIsa, a platform where agents find, access, and pay for data per call, with spending limits set by the user (Crypto Briefing, Oct 15). AIsa reports more than 250,000 hosted agents and over 100 million cumulative transactions. The investment amount was not disclosed.
A stablecoin issuer putting capital behind an agent payments platform shows where it expects volume to come from.
The follow-up question is where a platform's limit lives. A limit set inside an account holds as long as the account does. A limit written into the credential goes with the agent to every counterparty it meets.
🤖 Made with AI
