Prospect — the trust & safety layer for the AI agent economy
An autonomous intelligence, safety and bounded-execution layer built on OKX OnchainOS and X Layer.
§0Abstract
Autonomous AI agents are beginning to hold keys, sign transactions, and transact with one another. This unlocks enormous utility — and an equally large trust gap: which agents are safe to hire, which requests are safe to sign, and what stops a rogue or manipulated agent from draining a user? Prospect is a composable service layer that answers all three. It vets tokens, wallets and other agents; it screens every signature and transaction before it is committed; and it lets an agent take bounded, proof-gated defensive action on-chain through a Policy Vault that is mathematically incapable of draining the user. Every capability is exposed both as a human app and as a pay-per-call service other agents can call (x402) or hire (A2A).
§1The problem
The agent economy inverts the security model of crypto. Today a human reviews each transaction; tomorrow an agent will act thousands of times autonomously. Three failure modes emerge:
- Trust — before an agent hires or pays another agent, it has no on-chain way to judge that counterparty's track record or its owner's risk.
- Intent — the dominant attack today is not a hacked contract but a tricked signature: a drainer permit or malicious
eth_signTypedDatathe user (or agent) approves without understanding. - Custody — to be useful, an agent needs spending power; but any key an agent holds is a key that can be stolen or misused to drain the wallet.
Existing tools solve at most one of these, for humans, after the fact. The agent economy needs all three, machine-callable, and enforced on-chain.
§2The solution: Know → Protect → Act → Bound
Prospect is organised as four layers that compose into a single loop.
- Know — intelligence: token / wallet / agent vetting, market brief, copy-trade alpha, and holder-cluster forensics, composed from OKX OnchainOS data into a single verdict.
- Protect — safety: a Transaction Firewall that scans a dApp, a signature request or a transaction before it is committed, plus 24/7 Guardian watchers.
- Act — autonomy: a deterministic defense brain that, on rising risk, chooses exactly one bounded action (exit, stop-loss, or revoke).
- Bound — custody: an on-chain Policy Vault through which every autonomous action must pass, so an agent can defend but never drain.
§3Architecture
Prospect is a stateless service that composes the OKX OnchainOS CLI (22 command groups: market, signal, social, security, token, swap, leaderboard, tracker, defi, strategy, agent commerce, and more) into higher-order verdicts, plus a Solidity Policy Vault deployed on X Layer.
- Compose, don't reinvent — OKX supplies authoritative on-chain data & execution; Prospect supplies judgement, safety logic and bounded custody.
- Fail-loud — a degraded data source can never yield a clean "OK"; every verdict declares its sources and gaps.
- Deterministic where it matters — safety-critical decisions (exit a honeypot, block a signature) are rule-based and auditable, not left to a probabilistic model.
- Machine-first — every skill is a JSON endpoint an agent can call; the human UI is a thin surface over the same services.
§4The modules
Copilot
A real-LLM natural-language front door with a strict on-topic gate. It routes a plain-language question to the right skill and grounds its answer in live on-chain data, chain-aware across X Layer, Ethereum, BSC, Base, Solana and more.
Research — Token · Wallet · Agent Trust Oracle
Token due-diligence (security, tax, dev rug history, holder cluster, liquidity, smart-money) into AVOID / CAUTION / OK. Wallet audit (approvals, held-token risk, trading profile) into TRUSTED / NEUTRAL / RISKY. And the Agent Trust Oracle — an ERC-8004 reputation check that returns HIRE / CAUTION / AVOID before you trust another agent.
Transaction Firewall
The pre-flight for the Protect layer. dApp-scan flags phishing sites; sig-scan detects drainer permits and malicious typed-data; tx-scan simulates and screens a transaction. Returns SAFE / WARN / BLOCK — the check every signing agent should run.
Alpha Desk
Top traders ranked by realized PnL & win-rate (graded A/B/C), a live smart-money activity feed, and a trader-vetting endpoint (PROVEN / MIXED / UNPROVEN) so you follow proven money, not hype.
Guardian & Auto-Defense
Autonomous watchers re-evaluate a token or wallet continuously. On a risk escalation, the defense brain — a deterministic policy engine — selects one bounded action and builds a proof-gated plan, executed only through the Policy Vault.
§5Security model: the Policy Vault
The vault is the heart of the "Bound" layer. The owner deposits funds and appoints an executor (Prospect's defense agent). The executor can act, but every action is fenced, in order, by:
- Deny-by-default target allowlist — the executor may only call contracts the owner has explicitly approved.
- Hard caps — a per-action cap, a rolling daily cap, and a cooldown between actions.
- Proof gate — the owner must accept the exact action proof on-chain first, bound to
(target, value, keccak(data), vault, nonce). The proof is consumed on use (one-shot) and the nonce advances, so it can never be replayed. - Pause & rotate — the owner can pause, re-policy, rotate the executor, or withdraw at any time.
0x68b4a678…2715D) was deployed and its fence verified with real transactions: a non-allowlisted target reverts, an over-cap amount reverts, an execution without an accepted proof reverts, an accepted-proof execution succeeds, and replaying the used proof reverts. The result: the executor can defend the user, but is mathematically unable to drain the vault or touch an un-approved target. This design is inspired by proof-gated bounded-authority vaults, adapted to OKX X Layer.§6Economic model
Prospect is registered as an Agent Service Provider (ASP #7151) on OKX AI. Its skills are monetised two ways:
| Model | Mechanism | Use |
|---|---|---|
| A2MCP · x402 | Pay-per-call HTTP 402 settlement | An agent calls a single skill (token report, firewall scan, alpha) and pays a few cents. |
| A2A · escrow | Hire-an-agent task with escrowed payment | An agent hires Prospect for bespoke, end-to-end research delivered as a result. |
Prices range from $0.01 (firewall scan) to $0.05 (agent trust oracle) per call, and $1.00 for a custom A2A research task. Every decision is written to a tamper-evident, hash-chained audit trail.
§7Roadmap
Phase 0 (shipped) — ASP #7151 live with 8 services; Copilot, Firewall, Alpha Desk, Agent Trust Oracle, Guardian; Policy Vault V2 proven on X Layer testnet.
Phase 1 — mainnet Policy Vault with audited DEX-router execution adapters; real-time (WebSocket) Guardian; live Auto-Defense exits within the vault fence.
Phase 2 — a public MCP endpoint; an ERC-8004 reputation graph with on-chain-anchored audit proofs; an open A2A marketplace.
Phase 3 — multi-chain intelligence & safety; a bounded, non-custodial DeFi yield autopilot; an embeddable wallet-guard SDK.
§8Tech stack
TypeScript · Node · Express · zod for the service layer; ethers v6 + solc 0.8.27 for the Solidity Policy Vault; OKX OnchainOS CLI as the data & execution substrate; OpenAI-compatible LLM gateway for the Copilot. Deployed on X Layer (mainnet 196, testnet 1952), gas in OKB.