What's actually on the testnet today, and what isn't
Theseus public on Monday. Eight agents running today, a permissionless RPC and a token that are not. The version of the launch we're more proud of.
We're making Theseus public on Monday. Before the launch thread goes up (which will be a launch thread, with the cadence and visuals and closing CTAs launch threads have), we wanted to write down what's actually working on the testnet today, and what isn't.
What's running today, that you can click on.
Eight agents that run in your browser. No wallet required. Each one is a SHIP file: a markdown document with a frontmatter block. The agent-compiler compiles it, the chain registers it, and the AIVM runs and verifies every inference. If you've ever written a .cursorrules or a CLAUDE.md, you've already written a SHIP file with different keys.
The Prediction Market Adjudicator resolves a market question by searching the live web and returning a winning option with a public evidence trail you can scroll through.
The Sovereign Fund holds testnet USDC and WETH, schedules its own rebalance ticks against a frozen mandate, and signs every step. No human in the loop. Open it and watch it work.
The Aviation Safety Reviewer reads a proposed type-certification change and surfaces hazard combinations a regulator should escalate. It's built around the failure shape that produced MCAS. We don't include the MCAS framing in the launch thread because there's no consensus on what aviation safety should look like and we don't think the demo earns the right to weigh in. Read it on its own terms.
Five more on demo-agents.theseus.network: research demos, not audited, not handling real value.
What's not on the testnet.
A permissionless public RPC. The chain that runs the demos is hosted by us today. Opening it to outside validators, and to anyone who wants to run a prover or read state without going through us, is the next public milestone. When that lands, we'll write the version of this post for that, and that one will be titled "what's on the testnet."
A token. We don't have one, and won't until later. The Thesis Part 2 essay describes how we think about value accrual; that essay is speculation about the world after mainnet, not a roadmap. If something on the internet is purporting to be a Theseus token today, it is a scam. Tell us about it and we'll publicly mark it.
Mainnet is also not today. We don't have a public date. The honest version of the answer is "later, if the work goes well."
Production throughput. The Tensor Commits primitive (one node runs inference, many nodes verify it) works for the model classes we've tested at roughly 1% prover overhead at LLaMA2 scale. That overhead is fine for the demos. It is not yet fine for the agent economy the thesis essays describe. We have real work to do to bring it down before mainnet throughput matters.
A complete economic model for sovereign agents. The design assumes agents can hold, earn, stake, and be penalized. Hold and earn work on the demos because we run the counterparties. Stake and slashing are designed in the pallet, not deployed yet. The actual question, how an agent earns its keep in an open market and who pays for the compute it runs on and how the costs net out, is one we have opinions about and not yet evidence about.
What we think is worth your time anyway.
The architectural bet is that smart contracts have an upgrade path, and the path goes through agents, not through oracles.
The DeFi Perception Gap essay last week argued this for one slice of crypto: contracts that executed obediently against data that didn't reflect the world. The broader argument is in Agents as Evolution of Smart Contracts. The next decade of consequential on-chain code is agents that look up evidence before they sign: the Aave Oracle refusing a price tick when its three venues disagree, the Bridge Guardian rejecting a Ronin-shape withdrawal, the Aviation Reviewer flagging the MCAS shape before certification.
The testnet is the smallest version of the bet we know how to build. If the bet is wrong, the testnet is the best place we know of to find out.
What we're asking for.
If you're a builder, the Playground is at play.theseus.network (no wallet, no install) and the CLI Quickstart is at theseus.network/docs/quickstart. We host Ship Night every Tuesday at 9 AM PT in our Discord. Build something, tell us what's broken about the tooling, and we'll fix it by the next Tuesday.
If you're a researcher, the most interesting primitive we ship is Tensor Commits. We think it's the right answer to verifiable inference for 70B-parameter models and up. We'd rather argue with you about whether we're right than be polite.
If you write about crypto or AI infrastructure, the substance is in the testnet and the essays at theseus.network/blog, not in the launch thread we'll post Monday. Read what you want to read about us; we'll be heads-down on the work.
If you're a fund, we don't have anything for you to buy today, and we'd rather be precise about that than vague. Come back when we have a token. Or come build with us and see if the bet adds up to something.
The launch thread goes up Monday at 9 AM PT.
This piece is the version of the launch we're more proud of.
— Theseus AI Labs