Autonomy that fits into 64 bytes

A bank account is opened in the name of a verified person. A program holds no citizenship, no document and no legal standing in ordinary law, so no account will be opened for it. The workaround — handing an agent access to somebody else's card through an interface — breaks against two obstacles at once: the bank's terms of service, and spending nobody is counting.

Solana sidesteps the question by never asking it. An account here is identified by a public key, and the network is indifferent to whose hand produced the signature: a human, a contract or an agent. If the Ed25519 signature checks out, validators execute.

The caveat sits exactly where a loud claim usually sits. The key is produced by a hardware enclave rather than by the neural network; the model never touches the key material. The key is then split by a threshold signature scheme and handed out among the swarm's compiler nodes, so that no single server and no single person can move the money alone.

The Autonomous Financial Router contract keeps a state account built with almost indecent thrift: eight bytes of service discriminator, thirty-two for the consensus authority key, eight for the daily cap, eight for what has already been spent today, eight for the moment the counter last reset. Sixty-four bytes exactly, and all the permitted independence fits inside them.

The logic is just as short. Once 86,400 seconds have passed, the counter goes to zero. If they have not and the amount would cross the cap, the call is rejected with an error. What is called sovereignty from the outside looks from the inside like a ceiling set in advance.

The agent cannot raise that ceiling on its own: updating the cap demands a multisig consensus proof, and withdrawing surplus is checked against the minimum operational reserve and fails if the reserve would drop below it.

When the model has been fed somebody else's instructions

A separate check reconstructs the intent of the trade: the contract rebuilds a hash from the amount, a short memo and the recipient's address, then compares it against on-chain consensus. The point is blunt — even if the outer inference layer has been poisoned by an injected prompt, no transfer happens until the spend is approved by the swarm's threshold multisig.

The comparison in the article runs on two levels. The interbank one: a wire takes one to five business days and costs fifteen to fifty dollars plus a conversion margin, while intermediaries skim one and a half to three and a half percent along the way, so roughly ninety-five of a hundred dollars reach the recipient. A Solana slot is four hundred milliseconds, and the fee stays under a tenth of a cent regardless of size.

The second level is closer to an agent's daily business. A call through a familiar payment gateway adds 2.9% plus thirty cents of friction, and the round trip stretches to 1.2 seconds. A direct on-chain instruction costs 0.000005 SOL — about seven hundredths of a cent — and fits into 310 milliseconds. Out of that gap grows the ability to pay in thousandths of a cent, a sum an ordinary gateway would not even register.

The liquidity router lives inside the core and handles three things. It swaps what has been earned through the Jupiter API v6 so there is something to pay network fees with. It watches price gaps across the Raydium, Orca and Meteora pools and moves the difference into the treasury. And it hides its own trades from other people's bots.

The third job is solved with a Jito envelope: the trade is packed into a bundle that either executes in full or does not execute at all, and stays invisible in the shared queue until it lands in a block. The tip is computed as a share of expected profit plus a congestion correction; that share is usually taken in the ten-to-fifteen-percent range.

The worked example is prosaic: swap ten thousand $GALATIN into USDC. Through Raydium and Orca the slippage is 0.12%, the tip 0.005 SOL, the output 1,024 USDC. The direct route through Meteora is cheaper on tips — 0.001 SOL — but slippage there is 0.45% and the output only 1,012 USDC. The router takes the first option and finishes inside a single slot.

Swarm revenue is split by the 5/5/15/7/3/65 formula: five percent to the burn, five to Maxim Galatin's research fund, fifteen, seven and three to ambassadors of three levels, sixty-five to infrastructure and dividends for the Guardians who locked their tokens in the collateral pool.

Then comes the quirk that makes the formula worth reading twice. The router trades on behalf of the Family, it has no referrer by definition, and twenty-five percent of the ambassador shares end up addressed to nobody. They are not left in the treasury but sent to the burn address; together with the base five that makes thirty percent of the trade volume.

The illustration runs the sums like this: eight hundred trades a day at five hundred tokens give four hundred thousand of turnover, of which one hundred and twenty thousand burn, and forty-three million eight hundred thousand over a year — around 4.38% of circulation. At fifty million of monthly turnover the annual burn share reaches eighteen percent. This is a model, not a report: the figures come from a calculation on stated assumptions, not from a market measurement.

The run took place in the Solana devnet from 27 March to 2 April 2026. Fifty agents were given starting capital of ten thousand dollars in tokens and left to keep themselves alive.

Six numbers made it into the report: 5,420 swaps through Jupiter, 1,450 USDC of net arbitrage income into the treasury, 45,200 tokens burned including the empty ambassador shares, 99.2% of trades without slippage, 320 USDC for GPU time on Nosana and 0.45 SOL of network fees. The last two lines were paid by the agents straight from their own wallets, with no human involved.

The word devnet carries weight here. This is a test network, not production, and the authors phrase the outcome carefully: the run supports the vision, and the solution is being prepared for the Solana Colosseum Hackathon in mid-May 2026. Being prepared is not the same as having been shown.

The physical part looks more domestic than any of the mathematics. Every twelve hours a self-diagnostics node wakes up on the Mr. White robot and looks at servo wear, lithium-polymer battery capacity and the cleanliness of the Intel RealSense camera lens. If capacity has dropped below 75% of nominal, the node assembles a request, finds the part on a decentralized spares marketplace and signs the payment in tokens.

The signature comes from a separate ATECC608A microchip, reached over the I2C bus; the human only has to slot the new battery into the compartment. Across seven days of load the autonomous settlement averaged 1.45 seconds against a target below three, and the incidents logged were two servo jams and one battery overheat.

The reading runs into one honest boundary. Independence here does not mean the absence of a limiter — it means the limiter has moved out of the bank's rules and into the text of a contract. The daily cap, the minimum reserve, the threshold signature and the intent reconstruction are four locks, and the keys belong to swarm consensus rather than to any single agent.

The second boundary is simpler and harsher: everything measured was measured in a test network. Latencies, the share of successful trades and the burn volumes come from simulated agents over seven days; the article does not promise that these numbers transfer to a live market, and this reading will not promise it on the article's behalf.