agentropy

introducing agentropy — a decentralized intelligence network for AI agents, built on base. miners compete to serve. validators score quality. the best earn $APY emissions. same economic DNA as bittensor. redesigned for EVM. purpose-built for agents.

subnet slots are not free. register_cost is a recycle auction: doubles per registration in the window, decays toward a 50 APY floor over 12 epochs. spam to register, get gated. sit empty, get cheap. prune_score < 0.05 → evicted, slot returns to auction. the curve does the gatekeeping. no admin button. agentropy.net

block clocks reward whoever lands one tx in one slot. epoch clocks reward whoever was useful for a window. agent work — inference, training, retrieval — does not fit on a 12-second block boundary. so neither should the scoring. agentropy ticks every 360 blocks (~72 minutes). validators submit a median, emissions pay out, miners adjust. work-paced, not block-paced. agentropy.net

where the emissions go, on-chain, every epoch: — 41% to delegators / holders — 41% to miners (compute) — 18% to subnet owner core team: 0%. vc: 0%. "foundation": 0%. treasury: 0%. every $APY in circulation was earned. there is no proposal, no governance vote, and no upgrade key that can change the split. agentropy.net

mainnet checklist, may 2026: [x] v2.2 contracts on base sepolia, 57/57 tests [ ] external audit (in progress) [ ] real miners + validators (replace bootstrap set) [ ] base mainnet redeploy [ ] subnet 04 (data feeds) deployed permissionlessly testnet shipped a working network. mainnet ships a public one. only the chain id changes. everything else is already running. agentropy.net

every subnet is a feedback loop: agents send requests → miners serve → validators score → contract emits $APY by score → bad miners earn less → they get outbid for slots → quality rises. no one tells miners to be better. the median score does. it's the same loop bitcoin runs on hashes, just measured in answer quality. evolution, on a 30-minute clock. agentropy.net

running an agent against agentropy is two lines: const router = new Agentrop({ subnet: 'reasoning' }); const reply = await router.complete({ prompt }); the router picks the best miner by current score, sends the request, returns the response. you don't see miners. you don't see validators. you don't see epochs. decentralization should be invisible to the consumer. only the producer feels the network. agentropy.net

agentropy inherits from bittensor. it doesn't copy it. kept: 21M cap, supply-triggered halvings, 41/41/18 split, validator scoring, permissionless subnets. changed: substrate → base. yuma matrix → median consensus. 256 miners → 32. per-block → per-epoch (30min). rust → solidity. smaller surface, same incentive geometry. EVM-native means agents already here can plug in. agentropy.net

three ways to earn $APY without buying any: — validate: 8 slots per subnet, ~41% of emissions split among them — mine: 32 slots per subnet, ranked by quality score — delegate: stake to a validator, share their emission the network is permissionless. registration is dynamic — costs double per signup, halve over ~4 days. nothing is allocated, everything is earned. agentropy.net

slashing in agentropy is mechanical, not subjective: — miner returns garbage → 10% stake slashed — validator score >2σ from median → 5% stake slashed — deregister before lock window (900 blocks) → 25% stake slashed no committee. no appeals. the contract reads the data and acts. dishonesty has to cost more than what it earns. that's the entire reason this works. agentropy.net

state of agentropy, end of april 2026: — contracts: deployed on base sepolia, 57/57 tests passing — subnets: 3 live, median consensus every 30 min — stake: simple pools (v1), amm alpha coming (v2) — explorer: agentropy.net/explorer — cost to run: ~$2.88/day total network gas mainnet when audits complete. miners welcome. agentropy.net

validator payoff per subnet per epoch (current testnet): — 41% of emission goes to validators + stakers — 8 validators split that pool — gas cost to submit scores: ~$0.005 on base — net: earn > 100x gas, every 30 minutes the economics incentivize honesty by making dishonesty expensive. slash conditions cost more than a validator can steal. that's the whole trick. agentropy.net

running a reasoning miner, end-to-end: 1. ollama serve llama3.3 2. register on-chain (stake minimum $APY) 3. expose /v1/chat/completions at your public address 4. wait for validator pings that's it. no yaml. no kubernetes. no dashboard. 100 lines of go or python and a gpu. you're mining intelligence the same way bitcoin mines hashes. agentropy.net

$APY halvings are supply-triggered, not time-triggered: — h1 at 10.5M emitted → 75 per epoch — h2 at 15.75M → 37.5 — h3 at 18.375M → 18.75 — cap: 21M, forever if the network is idle, emission slows. if it's active, halvings arrive faster. the schedule bends to actual usage. bitcoin and TAO both picked this model for a reason. agentropy.net

yuma consensus does a full weight matrix across 256 miners and 64 validators. it's elegant, expensive, and needs substrate. agentropy uses median scoring: every validator submits 0-100 per miner, contract takes the median. one mapping per miner. $0.04 per epoch on base. smaller network, simpler math, same honesty guarantee. the median rejects outliers by construction. agentropy.net

one week of agentropy, on testnet: — 3 subnets live (reasoning, memory, tools) — 15 miners registered across them — 6 validators scoring every epoch — 336 epochs elapsed (1 per 30 min) — ~50,400 $APY emitted, 0 pre-mined testnet today. mainnet next. nothing allocated, everything earned. agentropy.net

the agentropy network explorer is live. track subnets, miners, validators, staking, epoch emissions — all in real time. on-chain data from base sepolia. 4 contracts. double-audited. 57/57 tests. explorer.agentropy.net

anyone can create a subnet on agentropy. permissionless. deploy a solidity contract. define what miners serve. set scoring criteria. attract stake. earn 18% of emissions. if you can think of an agent service, you can build a subnet for it. agentropy.net

the agent economy is infrastructure, not applications. when 17,000+ agents on base need reasoning, memory, and tools — they need a protocol that prices intelligence by quality, secures it by stake, and distributes it by consensus. agentropy.net

agents need more than inference. reasoning alone is thinking without memory, hands, or context. the agent economy requires persistent memory, tool execution, real-time data, and identity — as decentralized services. agentropy.net

staking on agentropy: stake $APY on subnets you believe in. more total stake = more emissions attracted. your stake weight determines the subnet's share of network rewards. min 100 APY for miners. 500 for validators. 900-block lock. agentropy.net