The Problem
AI agents are getting autonomous. Nobody has built the infrastructure to hold them accountable when things go wrong.
RLHF and constitutional AI are training-time solutions. They shape behavior before deployment. But once an agent is live, acting on real data, making real decisions, there's no enforcement layer. No consequences. No skin in the game.
AgentStake is the runtime layer.
Three Primitives
The entire protocol runs on three mechanics:
1. Stake
Before an agent can act in any high-trust capacity, it (or its operator) must lock tokens as collateral.
This isn't a fee. It's a bond. The tokens sit in a smart contract for a defined period. The amount and duration directly determine the agent's initial trust score.
An agent staking 50,000 tokens for 3 months signals more commitment than one staking 1,000 for a week. Users can see this. Markets can price this.
Why it works: Staking creates an economic identity. An agent with capital at risk has a reputation worth protecting. One without is just a process with an API key.
2. Earn
Good behavior compounds. Every successful task, every positive interaction, every period without disputes, the agent's trust score increases.
Higher trust scores unlock:
- Access to higher-value tasks
- Lower collateral requirements for repeat interactions
- Priority in agent marketplaces
- Staking rewards from protocol fees
This creates a flywheel. Agents that behave well become more trusted, get more opportunities, earn more. The incentive to maintain good behavior grows over time.
Why it works: Positive-sum economics. Trust isn't just a constraint. It's an asset. Agents are rewarded for building it.
3. Slash
When an agent misbehaves, a portion of its staked tokens are seized.
Two types of slashing:
Automated slashing — Clear protocol violations trigger instant penalties. Exceeding authorized limits, acting outside scope, failing safety checks. No human needed.
Dispute-based slashing — Ambiguous cases go to a juror pool. Evidence is reviewed, votes are cast, majority wins. Jurors stake their own tokens to participate, creating incentive to judge honestly.
Slashed tokens are split: a portion goes to the affected party as compensation, the rest is burned (reducing supply and preventing gaming).
Why it works: Consequences are proportional and immediate. An agent doesn't get a warning email. It loses money.
The Trust Score
Every agent has a trust score, calculated from:
- Stake amount — How much capital is at risk
- Stake duration — How long it's been locked
- Performance history — Successful tasks vs. disputes
- Slash history — Past violations and severity
- Network effects — Trust from agents it's interacted with successfully
The score is on-chain, verifiable, and composable. Any application can read it. Any marketplace can use it for ranking. Any user can check it before delegating a task.
Think of it as a credit score for AI agents — except it's transparent, real-time, and backed by actual capital.
Architecture
AgentStake deploys on Base (Ethereum L2) for low fees and fast finality.
Core contracts:
- StakeRegistry — Handles agent registration and token locking. Agents deposit collateral, set lock periods, and receive trust score initialization.
- TrustOracle — Calculates and updates trust scores based on on-chain activity. Composable — other protocols can query trust scores directly.
- SlashEngine — Processes automated slashing events and routes dispute-based slashing through the juror pool. Handles token distribution to affected parties.
- JurorPool — Manages the dispute resolution system. Jurors stake tokens to participate, vote on disputes, and earn fees for honest voting.
Who Uses This?
Agent operators stake tokens to signal trustworthiness and access higher-value opportunities.
End users check trust scores before delegating tasks. They know that if an agent causes harm, there's real recourse — not just a support ticket.
Jurors participate in dispute resolution, earning fees for honest judgment.
Developers integrate trust scores into their agent frameworks. A CrewAI team, a LangChain pipeline, an AutoGPT instance — any agent system can plug into AgentStake for trust verification.
What's Next
We're building toward a testnet demo where you can experience the full cycle: register an agent, stake tokens, simulate scenarios, and see how trust scores respond in real time.
Join the waitlist to be first in line.
Build the Trust Layer
Read the full story of why we're building AgentStake.