Agents can act, coordinate and pay — but never beyond a defined budget, permission, risk or verification boundary.
The missing layer in agent infrastructure is not another model or marketplace. It is a policy-controlled execution layer that can decide whether an agent is still allowed to act — at every step — and leave evidence when it does.
Intent → Policy → Risk → Execute → Verify → Recover → Receipt
The ETHOnline 2026 Continuity submission: live blockchain data informs an autonomous buyer, bounded authority controls the spend, and an independent verifier decides whether USDC is released.
The Graph → Policy Engine → Privy authorization → Arc USDC escrow
→ Seller Agent → Independent Verifier → VERIFIED / BLOCKED / FROZEN
The repository includes a judge-first Demo, Continuity before/after record, machine-readable evidence manifest, sponsor fit notes and reproducible checks.
- Agent Control Plane — generic policy-controlled execution foundation.
- Monad AgentGuard — Monad deployment, benchmarks and baseline evidence.
- Binance AgentGuard — Binance Agentic execution adapter with independent verification.
These are one capability expressed through different chains and execution venues, not unrelated products.
| Layer | What it controls |
|---|---|
| Policy | permissions, budgets, targets, approvals and time windows |
| Risk | exposure, duplication, slippage, runtime drift and circuit breakers |
| Execution | exchange, chain, payment, MCP, x402 and workflow adapters |
| Verification | outcome checks before release or settlement |
| Recovery | cancel, refund, freeze, retry and escalation paths |
| Receipts | auditable decisions, proofs and execution history |
Earlier repositories across Arc, Solana, Arbitrum, OKX, BNB, DeFi and privacy are preserved as experiments and continuity context. The current product thesis is AgentGuard: make autonomous economic activity bounded, verifiable and recoverable.
- agentpay — programmable wallets and payment rails.
- agentbank — multi-agent treasury automation.
- agentcourt — proof and dispute handling.
- kitehive — Agent-to-Agent labor-market experiment.