MetaMask launched Agent Wallet this week, and the pitch is blunt: let an AI agent trade crypto on its own, inside limits a human sets first. Decrypt reported the wallet supports swaps, perpetuals, prediction-market bets, and staking across Ethereum-compatible chains plus Hyperliquid — which is a wider surface than most "agent wallet" launches bother covering. This isn't MetaMask's idea alone; Coinbase shipped agentic wallets back in February and MoonPay has one in development. But MetaMask controls roughly a quarter of the wallet market, per the same report, so its version is the one most builders will actually run into.
The interesting part isn't that a wallet vendor wants agents connected — every wallet wants that traffic. It's how MetaMask drew the permission boundary, because that boundary is the whole ballgame once an agent can move funds without asking first.
Two modes, two different bets on trust
Agent Wallet ships with two settings:
- ▸Guard Mode (default) — the agent can act inside an allowlist of pre-approved protocols and spending limits, but anything outside that allowlist needs a 2FA approval from the human.
- ▸Beast Mode — drops the allowlist entirely. The agent trades freely within a spending cap, no per-transaction sign-off.
Spending caps apply in both modes. The names are doing some marketing work — "Beast Mode" sells autonomy, "Guard Mode" sells safety — but the actual design choice underneath is a familiar one: allowlist plus human-in-the-loop versus cap plus full delegation. That's the same tradeoff every agent-wallet builder has to make, MetaMask just shipped both options instead of picking one.
What's claimed vs what's shown
MetaMask says every transaction gets simulation, threat scanning, and MEV protection, and that eligible trades carry up to $10,000 a month in loss coverage. Those are real, specific commitments — a loss-coverage number is more concrete than most "we take security seriously" copy. But they're also self-reported at launch, with no outside security review or incident history to check them against yet. "Simulated and scanned" is a claim about process; whether it actually catches the failure modes that matter — a malicious contract, a drained allowance, a bad price feed — is something only a track record proves. Worth remembering: even in Beast Mode, MetaMask keeps a 2FA backstop for transactions it flags as malicious, so the "remove the allowlist" pitch isn't actually removing every check.
One detail that's easy to miss and genuinely useful: gas gets settled in whatever token the agent is already moving, so an agent doesn't need to separately hold and manage native gas tokens across every chain it touches. That's a small thing that quietly removes a real operational headache for anyone running an agent across multiple chains.
There's no token attached to Agent Wallet. In a category where half the launches come with a ticker, that's worth noting on its own.
Who else is already here
MetaMask lists Claude Code, Codex, Cursor, and OpenClaw as compatible AI tools at launch — the wallet is positioning itself as infrastructure those coding/agent environments plug into, not a standalone chatbot with a wallet bolted on. That's the right instinct for an agent-economy wallet: agents are increasingly built inside a harness, not summoned from scratch, so the wallet needs to meet builders where the harness already is.
The Sato read
Agent Wallet is a permissions product wearing a trading-features press release. Swaps, perps, prediction markets, and staking are the surface; Guard Mode vs Beast Mode is the actual product. That's the right place for a wallet vendor to compete — not "can an agent trade," every wallet can wire that up, but "what happens the one time the agent is wrong." A spending cap and a loss-coverage number are the beginning of an answer, not the end of one.
If you're building an agent that needs to hold and move funds, the checklist this launch reinforces doesn't change: spend limits, an allowlist for anything outside routine activity, and simulation before signing. Whether MetaMask's version of that checklist holds up is a question for six months of incidents, not a launch post.
What to watch
Watch for the first disputed loss-coverage claim, or the first Beast Mode drain that tests whether the malicious-transaction backstop actually catches what it's supposed to. Both will tell you more about Agent Wallet than the launch copy does. And watch whether Coinbase and MoonPay converge on the same two-mode pattern — if the whole category lands on "allowlist mode" plus "capped autonomy mode," that's the industry quietly agreeing on the only two designs that survive contact with an agent that's occasionally wrong.
Sources
- ▸[Morning Minute: MetaMask Hands AI Agents a Wallet](https://decrypt.co/375093/morning-minute-metamask-hands-ai-agents-a-wallet) — Decrypt