DEV Community

AgentWallex
AgentWallex

Posted on

The $30M Question Nobody's Answering Right

#ai #payments #webdev #crypto

VCs just deployed $65M across two "AI-native" payment companies. Catena Labs raised $30M for an AI-native bank. Sapiom raised $35M for agent cost control infrastructure. Circle's co-founder is filing for a national trust bank charter.

Everyone's building the banking layer.

Nobody's building the control layer.

And that's where the risk actually lives.

The Wrong Question

Newsweek, CryptoSlate, and Audacy are all running the same headline variants: "Who pays for AI agent mistakes?" They're asking about liability, insurance, legal frameworks.

They're asking the wrong question.

The right question: "How do I architect payments so the mistake can't happen?"

This isn't semantic hairsplitting. It's the difference between reactive legal cleanup and proactive system design. Between hoping your insurance covers it and making sure there's nothing to cover.

Policy Engines as Infrastructure

Here's a concrete scenario:

Your AI procurement agent has a wallet. It's authorized to buy SaaS subscriptions, API credits, cloud compute. But you don't want it spending $50K on OpenAI credits at 3am because a loop went wrong.

Traditional approach: human-in-the-loop approval workflows. Every transaction waits for a Slack notification and a thumbs-up. Latency added. Autonomy removed. Still doesn't prevent the agent from requesting the bad transaction.

Policy engine approach: hard limits enforced at infrastructure level.

{
  "agent_id": "procurement_agent_001",
  "daily_spend_cap": 500,
  "allowed_recipients": [
    "openai.com",
    "anthropic.com",
    "stripe.com"
  ],
  "rate_limit": {
    "max_transactions_per_hour": 10
  },
  "time_rules": {
    "block_weekends": true,
    "block_hours": [22, 6]
  }
}
Enter fullscreen mode Exit fullscreen mode

This agent literally cannot:

  • Spend more than $500 in a day
  • Pay anyone not on the allowlist
  • Execute more than 10 transactions per hour
  • Transact on weekends or between 10pm–6am

Not "shouldn't." Cannot. The authorization fails at the protocol level before funds move.

Fast Doesn't Mean Reckless

The counterargument: "Doesn't this add latency?"

No. Policy checks run in <150ms. Same authorization call that verifies MPC signatures also verifies policy compliance. The agent gets a yes/no answer in the time it takes to ping an API.

Compare that to human-in-the-loop workflows:

  • Slack notification sent
  • Human checks phone (maybe)
  • Human evaluates context (maybe correctly)
  • Human approves (minutes to hours later)

AgentWallex policy checks happen faster than a card authorization at a grocery store. And unlike humans, they don't get tired, distracted, or approve things they shouldn't.

The CFO's Best Friend

Enterprise AI teams care about three things:

  1. Auditability: Every transaction has a policy evaluation in the audit log. "Why did this payment go through?" has a deterministic answer.

  2. Explainability: Policy rules are readable JSON, not black-box ML models. Your compliance team can review them.

  3. Insurability: Insurance underwriters love deterministic controls. "We have per-agent spend caps and vendor allowlists" beats "we have really smart people monitoring things."

CFOs don't want to answer "who's liable?" after a $100K mistake. They want to show the board that the mistake was architecturally impossible.

Where VCs Are Betting

Catena and Sapiom raised on the thesis that agents need banking infrastructure. They're right — agents do need accounts, custody, settlement rails.

But the bank account is where money sits. The control layer is where money moves.

You can have the most sophisticated AI-native bank in the world. If the agent has carte blanche to withdraw from it, you haven't solved risk. You've just made risk faster and better-capitalized.

Policy engines sit between the agent's intent and the bank's execution. That's the choke point. That's where you enforce rules, check limits, block bad actors, generate audit trails.

MPC + Policy = The Full Answer

AgentWallex combines two things:

MPC wallets: 2-of-3 threshold signing. The agent never holds private keys. It can authorize payments, but it can't unilaterally move funds. Even if the agent's server is compromised, the attacker doesn't get the keys.

Policy engine: Every authorization request hits the policy layer before MPC signing happens. Even if someone compromised the agent logic and tried to authorize a $1M payment, the policy engine rejects it before keys are ever involved.

This is defense in depth. Not "MPC or policy." Both.

The Headlines Will Keep Coming

"Who's liable when AI agents make mistakes?" will be asked a thousand more times. Legal teams will draft frameworks. Insurance companies will write policies. Regulators will publish guidance.

All necessary. None of it solves the problem.

Because liability is a question you ask after the mistake happens.

Policy engines are how you make sure the mistake doesn't happen.

VCs are betting tens of millions on the banking layer. That's the foundation. But the control layer is the walls, the locks, the alarms.

You need both.

And right now, everyone's building the foundation and hoping the walls take care of themselves.

AgentWallex: The payment gateway for AI agents. MPC-secured wallets. Policy-enforced limits. <150ms authorization. No human in the loop.

Sandbox live at app.agentwallex.com — 3,600+ teams already building.

Follow & Try AgentWallex

Top comments (0)