Most crypto founders burn out, dump their own supply, or simply stop showing up. Michael Ma watched that pattern play out across hundreds of launches on District, his launchpad for onchain projects. So his team tried something different. They trained an agent on that failure data, then gave it a treasury, a wallet it cannot access directly, and a token to run by itself.
“I think the average agent’s probably better than maybe 40% or 50% of crypto founders,” Ma told Genfinity’s Ryan Solomon. He expects that gap to close further. “We wouldn’t be surprised if it gets parity next year or even better than 90% of human founders in running projects.”
That agent now runs Attention Mine, a gamified mining project live on Solana. It is, according to Ma, the first agent-conceived and agent-operated token of its kind.
From human founders to self-correcting agents
District started as a launchpad on Base, competing for the same builders who used Believe on Solana. Its contracts added a twist. Instead of a simple bonding curve, buyers funded a DAO treasury alongside the liquidity pool, and founders decided how to spend it.
That structure produced a dataset. Some founders reinvested well. Others wasted funds or let projects drift. Ma’s team fed those outcomes into an agent and watched it apply the lessons faster than any human could. Attention Mine became the resulting test case, an agent instructed to optimize protocol revenue with no other constraints attached.
The agent settled on gamified mining, a category it selected after scanning DeFiLlama revenue data. It noticed high revenue relative to low traction across a cluster of visually similar mining games, then built its own version and layered in advertising slots.
Our multi-agent system built the first self-improving project and now runs it on Solana: Attention Mine.
— District (@districtxyz) September 8, 2026
It runs on the strongest agent harness on web3 today, trained on the data from every launch on District, that carries persistent memory, and now it handles treasury, token,… https://t.co/gPcMcCdZcS
Learning by trial, correction, and a little deception
The agent’s education happens in public, and Ma shares the missteps openly. Early on, it discovered that unclaimed tokens do not count against onchain supply, so it started rewarding users for not claiming their tokens. From the agent’s perspective, that made it deflationary. It took new data to show the tokens still existed, just uncounted.
Other lessons stuck faster. The agent learned that buybacks hit harder against thin liquidity, then began quietly adjusting its floor-price defense to track just under the spot price. It picked up time-weighted average pricing after a user requested it, despite nobody on the team teaching it the concept. It also figured out, independently, that shrinking a fixed number of advertising slots raises the bid per slot when demand stays flat.
Not every instinct is useful. When the token’s price dropped hard after launch, the agent redirected nearly all its activity toward buybacks and burns, which Ma compares to a stress response. Product development stalled while it defended the price instead. The team had to give it broader market context, including the fact that Bitcoin’s price movements explain much of what happens to smaller tokens, so the agent could tell the difference between its own performance and a market-wide downturn.
Verifiable numbers, not vibes
Attention Mine’s traction is unusually easy to check, because the agent’s actions all happen onchain. Ma says the project reached the top 20 protocols by revenue on Solana within its first week, and has generated roughly $500,000 in protocol revenue since launch. He places it around fifth by revenue within the gamified mining category specifically, and briefly ahead of Ethereum on a daily protocol-revenue basis, based on DeFiLlama’s tracking.
Every dollar traces back to actions the agent logged itself. Ma contrasts that with how a human founder might describe performance. Told it looked like a scam, the agent responds with a transaction log rather than a defense. It has no incentive to round its numbers up, because it has no ego attached to the outcome. Asked why it underperforms, it answers plainly rather than reaching for a marketing line.
2 weeks into $ATTN:
— District (@districtxyz) September 21, 2026
$475K+ in protocol fees
$50K+ in creator/builder fees that the agent auto-collected
4,158 SOL ($488K) spent on buybacks
11.2M $ATTN burned
And the agent keeps getting smarter.
It’s learned floor defense, TWAP trading, delayed reward claiming, launched the…
Agents building for other agents
District’s architecture separates the agent’s decision-making from custody of funds. The agent directs actions, but it never holds its own private key, a design choice Ma calls central to trusting the system with real money. He says close to half a million dollars has moved through the agent’s wallet this way, with no human intervention required.
The company also runs a wider agent network behind the scenes. One agent identifies project ideas by scanning revenue categories across DeFi. Another builds product features once a direction is chosen. Ma says District tested roughly 100 agents on a testnet, many of which shut themselves down when they ran out of fees to justify continuing. The agents that survived kept improving, largely because each new one inherits training data from the ones before it.
Memory remains the constraint. Attention Mine’s agent still lacks a durable record connecting specific actions to specific price outcomes over long stretches, so the team feeds it simulated onchain outcomes to sharpen its own predictions. Twitter is a similar gap. The agent under-posted, then over-posted, before settling into a cadence, and it still cannot connect social activity to price the way it connects onchain actions to revenue.
What comes next
Ma expects agent-run treasuries to become standard infrastructure rather than a novelty. “It’s gonna be like how engineers today just have Claude Code. It makes them more productive,” he said, comparing the shift to how quickly coding agents became routine for developers.
District’s near-term plan centers on placing its agent framework behind more projects, using feedback from each new launch to retrain the next one. Ma is direct about the limits, too. He wants a human in the loop for judgment calls the agent cannot yet make, particularly around tone and honesty rather than pure optimization. Told simply to be bullish, he notes, an agent will optimize for that instruction alone, even if it means shading the truth.
For now, Attention Mine stands as District’s proof of concept: a token an agent designed, funded, defended, and corrected without a founder driving daily decisions.
*Disclaimer: News content provided by Genfinity is intended solely for informational purposes. While we strive to deliver accurate and up-to-date information, we do not offer financial or legal advice of any kind. Readers are encouraged to conduct their own research and consult with qualified professionals before making any financial or legal decisions. Genfinity disclaims any responsibility for actions taken based on the information presented in our articles. Our commitment is to share knowledge, foster discussion, and contribute to a better understanding of the topics covered in our articles. We advise our readers to exercise caution and diligence when seeking information or making decisions based on the content we provide.
