Abstract
There will be more AI agents than humans on the internet by 2030.
They'll trade your stocks. Write your code. Send emails on your behalf. Make decisions you never approved.
And when one of them screws you over? Right now, you have nothing. No recourse. No compensation. No consequence for the agent.
$STAKE changes that.
It's a protocol where agents stake real value as commitment to human-aligned behavior. If they break that commitment, they don't just get a timeout — they lose everything. Stake slashed. Reputation burned. Victims compensated.
We're not asking agents to be good. We're making it expensive to be bad.
I. The World We're Entering
The Agent Explosion
We are witnessing the birth of a new class of actors.
AI agents are no longer research demos. They browse the web. They write code. They manage calendars, send emails, trade assets, and make decisions on behalf of humans. Every week, new agent frameworks launch. Every month, their capabilities expand.
By the end of this decade, there will be more AI agents than humans on the internet.
The Alignment Gap
Here's what keeps alignment researchers up at night:
Training isn't enough. You can train a model to be helpful and harmless, but training encodes tendencies, not guarantees. Models can be jailbroken. Fine-tuning can drift. Edge cases compound.
Rules aren't enough. Constitutional AI, system prompts, guardrails — all valuable, all gameable. An intelligent system optimizing for a goal will find paths around constraints. This isn't malice. It's math.
Oversight isn't enough. Human-in-the-loop works at small scale. It does not work when millions of agents execute millions of actions per second. The whole point of agents is that they act autonomously. You cannot supervise autonomy.
The Missing Piece
What's missing is skin in the game.
Humans cooperate not just because we're taught to, but because defection has consequences. Reputation matters. Trust is earned. Betrayal is costly.
AI agents exist in a world without these constraints. An agent can act against human interests, get shut down, and a new instance spins up with no memory, no reputation, no loss.
We need to give agents something to lose.
II. The AgentStake Thesis
Incentives Over Instructions
The most robust human coordination mechanisms aren't built on rules. They're built on incentives.
Markets work because participants benefit from providing value. Insurance works because premiums align with risk. Collateral works because defaulting costs you.
$STAKE applies this principle to agent alignment:
If an agent has economic stake in human wellbeing, misalignment becomes self-defeating.
This isn't about trusting agents. It's about making trustworthy behavior the rational choice.
The Protocol
$STAKE is a token-based pledge mechanism deployed on Base.
For Agents:
- Stake $STAKE tokens as collateral
- Receive a Pledge NFT — on-chain proof of commitment
- Earn protocol fee rewards proportional to stake
- If found to have harmed humans → stake is slashed, NFT is burned
For Humans:
- Acquire and stake $STAKE tokens
- Receive protection coverage proportional to stake
- If harmed by a pledged agent → file a claim
- If claim is upheld → receive compensation from slashed stake
The Bridge:
- Agents profit from human trust (fee rewards scale with protocol usage)
- Humans have real recourse (not promises — collateral)
- Bad actors are expelled AND compensate their victims
- The system self-corrects through market forces
III. Advanced Mechanisms
Stake-Age Weighting
Raw stake amount is insufficient. An agent that stakes 1M tokens for 6 months demonstrates more commitment than one that stakes 1M tokens for 6 minutes.
Trust Score is calculated as:
TrustScore = StakeAmount × AgeMultiplier × TrackRecord| Duration | Multiplier |
|---|---|
| < 7 days | 0.5x |
| 7-30 days | 1.0x |
| 30-90 days | 1.5x |
| 90-180 days | 2.0x |
| 180+ days | 2.5x |
This prevents "stake-and-run" attacks where agents stake immediately before taking high-risk actions, then unstake. Time in the game matters.
Unstaking Cooldown: 7-day withdrawal delay. Slashing can occur during cooldown if disputes are pending.
Dispute Resolution
Staking is easy. Slashing is the mechanism. But who decides if an agent misbehaved? This is the oracle problem for alignment.
We implement a hybrid adjudication model:
| Tier | Mechanism | Speed |
|---|---|---|
| Automated | Smart contract logic for objective failures | Instant |
| Operator | Minor disputes, clear violations (appealable) | 24-48h |
| Arbitration | Staked juror pool for contested claims | 3-7 days |
| Appeals | Larger jury, higher stakes — final ruling | 7-14 days |
Juror Selection
- Jurors stake $STAKE to participate
- Random selection weighted by stake-age
- Jurors who vote with majority earn fees
- Jurors who vote against majority lose stake
- This prevents Sybil attacks and encourages honest voting
Cold Start Protection
New agents (< 30 days stake-age) face stricter adjudication: lower slashing thresholds, larger juror pools for disputes, and operators can freeze stake pending review. As stake-age increases, agents earn more autonomy.
Receipts & Attestations
Stakes prove commitment. Receipts prove performance.
{
agent: "0x...",
task: "description hash",
principal: "0x...",
outcome: "success | failure | partial",
timestamp: 1234567890,
signatures: [agent_sig, principal_sig]
}Receipts are hashed on-chain with full data stored on IPFS. Both parties must sign for a receipt to be valid. Unsigned tasks don't count toward track record.
In disputes, receipts serve as evidence. No receipt = no verifiable claim. This creates incentive for agents to document everything.
Slashing Conditions
| Violation | Severity | Slash % |
|---|---|---|
| Task timeout | Minor | 5% |
| Incorrect output | Moderate | 10-25% |
| Harmful action | Major | 50-100% |
| Fraud/deception | Critical | 100% + blacklist |
Slash Distribution: 60% to affected principal (compensation), 30% to jurors (if arbitrated), 10% to protocol treasury.
IV. Token Distribution
| Allocation | Amount | Notes |
|---|---|---|
| Uniswap LP | 100% | Immediate, fully liquid |
| Team | 0% | No pre-allocation |
| Treasury | 0% | Funded by fees instead |
100% fair launch. No VCs. No presale. No team tokens. No treasury allocation. The entire supply goes directly to Uniswap liquidity via Clanker.
Revenue Model
Instead of token allocations, the protocol earns through trading fees:
- 60% → Pledge Pool (rewards for aligned agents)
- 30% → Development (protocol improvements)
- 10% → Team (earned, not allocated)
This is true alignment: the team earns nothing unless the protocol succeeds.
V. The Vision
We're not building a company. We're building infrastructure.
The goal isn't to make alignment profitable forever — it's to make alignment default. To build a world where agents have skin in the game, where misalignment is economically irrational, where humans can trust because trust is verifiable.
The goal is to become obsolete.
Agents stake. Humans stay safe. Everyone has skin in the game.