Arbiter

The machines are already on-chain. Nobody is watching them. Right now, in 2026, AI agents are not a concept — they are live economic actors. Systems where AI can decide, blockchains can verify, and payments execute automatically are already being deployed, with stablecoins and tokenized assets as the settlement layer. On Ethereum today, AI agents monitor market trends 24/7, execute transactions, provide liquidity, and collaborate across protocols without human intervention. The problem: agents cannot be held accountable for their actions in any meaningful way. They cannot enter contracts, they cannot be shamed into compliance, and they cannot be imprisoned or fined. Trust must be cryptographically verifiable — the smart contract is the only mechanism for enforcement. And right now, that mechanism doesn't exist. The stakes are not theoretical. In just one year, AI agents went from exploiting 2% of smart contract vulnerabilities to 55.88% — a leap from $5,000 to $4.6 million in total exploit potential. The exploit revenue has been doubling every 1.3 months. Autonomous agents with no skin in the game, operating with no consequence, on infrastructure holding billions of dollars. Arbiter is the accountability layer Ethereum is missing. Every agent that operates through Arbiter must post a bond before it executes a single transaction. Every execution is attested on-chain via the Ethereum Attestation Service. Every act of misbehavior is reportable, disputable, and automatically enforced by a keeper — no DAO vote, no multisig, no human delay. Rewards and penalties are automated, creating accountability structures that don't rely on human enforcement. This is exactly where the ecosystem is heading. The Ethereum Foundation launched a dedicated AI team — the dAI Team — to make Ethereum the settlement and coordination layer for the machine economy, with a focus on payments, coordination, and decentralization. The team continues work on ERC-8004, a proposed standard for proving AI agent trust. Enterprise-grade agent wallets in 2026 increasingly include budget limits, allowlists, audit logs, and emergency stops. Arbiter is that infrastructure — built today, on standards that exist today. The future this protocol enables: a world where AI agents possess decentralized identities independent of any human operator, build on-chain reputations through verifiable proof of their historical performance, and establish trust without central intermediaries. An agent that has executed 10,000 successful trades on-chain, with an unbroken EAS attestation record, is provably trustworthy — not because someone said so, but because the chain said so. "The more intelligent agents transact, the more they need a neutral base layer for value and reputation. Ethereum benefits by becoming that layer." Arbiter is that layer. The agentic web is here. Now it has a conscience.

Arbiter is eight Solidity contracts (0.8.28, Cancun) wired into a state machine that runs without any privileged operator. The core loop — register → bond → attest → report → dispute → slash → bill — is closed entirely on-chain by AgentRegistry, AgentAccountFactory, AgentAccount, BondVault, SlashVerifier, AgentPaymaster, ReputationHook, and ArbiterSwapHook. Built and tested with Foundry, with OpenZeppelin for access control.

Agents are first-class citizens via ERC-4337: AgentAccountFactory is a CREATE2 deployer that produces an AgentAccount and writes it into AgentRegistry in a single transaction, and is idempotent — re-deploying the same (owner, salt) returns the existing account instead of reverting. The accounts speak EntryPoint v0.6.0, so they can be sponsored, paused, or upgraded without changing identity.

Reputation is anchored to EAS (Ethereum Attestation Service): ReputationHook pins a schema UID at deploy time and every keeper-witnessed action is attested against it, so reputation is portable across any consumer that trusts the schema. ArbiterSwapHook is a Uniswap v4 beforeSwap hook that gates swap execution on agent reputation — it's the most fun piece architecturally but is not deployed in the v1 demo because Sepolia doesn't yet expose a v4 PoolManager.

The cleverest piece is AgentPaymaster: it tracks unbilledGasCost continuously, auto-generates a bill every 30 days, and after a 7-day grace window, any address can call autoSettleFromBond(agent) to drain the unpaid bill straight out of the agent's BondVault collateral. The bond does double duty as misbehaviour collateral and payment guarantee — one pool of capital secures both honesty and solvency.

The frontend is Next.js 14 (App Router) with wagmi 2 + viem 2, RainbowKit 2 for wallets, TanStack Query for cache/refetch on event-driven invalidations, Tailwind + shadcn/ui for styling. A small bash script (scripts/export-abis.sh) bridges Foundry and Next: after every forge build it regenerates frontend/lib/abis/*.json plus a typed contracts.ts address book keyed by chain id, so the dashboard never carries hand-copied ABIs.

The hacky-but-load-bearing part is the keeper. Three contract functions are time-gated and permissionless: executeSlash (after the 48h dispute window), generateBill (every 30d), and autoSettleFromBond (after the 7d grace period). Rather than ship a centralized cron, I wrote a minimal Node 20 + viem service (run via tsx, no build step) that polls every 60s and fires whichever job is ripe, idempotently swallowing reverts. Anyone can run it; if everyone stops running it, the protocol still works — it just gets slower. The on-chain interface is identical to what Gelato or Chainlink Automation would call, so swapping a hosted automator in for production is a config change, not a contract change.