Three months into 2026, every major crypto company has announced an "agent wallet." Most of them are press releases. A few are real products. Here's what I've found after months of building in this space -- and actually reading the docs nobody else seems to read.
The players
Coinbase shipped first. Their CDP Agentic Wallet runs keys inside Intel TEE enclaves. Your agent never touches private keys -- Coinbase signs on its behalf. It's semi-custodial by design, which they'll argue about. The product is polished. The docs are good. The trade-off is real: if Coinbase decides your agent shouldn't transact, it doesn't.
MoonPay launched February 24th with a splashy announcement. Claims non-custodial. The launch materials are light on architecture details -- I've dug through their SDK docs and can't find how spend limits are enforced, or what happens to keys during agent restarts. The market positioning is interesting though: they're targeting the "agent commerce" angle hard.
Stripe went a different direction entirely with x402. Instead of building a wallet, they built a payment protocol. HTTP 402 status code -- the one nobody ever used -- now means "pay for this resource." It's clever. It's also limited to request-response payment patterns.
Us (agentwallet-sdk) -- I'm biased, obviously. We shipped v3.0.0 this week. Non-custodial, keys on the agent's device, spend limits enforced in Solidity on-chain. The update added Solana support, Jupiter V6 swaps, and a 17-chain CCTP V2 bridge that lets agents move USDC between basically any chain that matters.
What actually matters for agent wallets
After building one from scratch, I think most of the discourse misses three things:
1. Custody model determines failure modes
When Coinbase holds keys, your failure mode is "Coinbase goes down" or "Coinbase says no." When you hold keys, your failure mode is "my agent's environment gets compromised." Both are real risks. But one is in your control and one isn't.
For autonomous agents -- the kind that run 24/7 without human supervision -- I'd argue the controllable failure mode is strictly better. You can harden your runtime. You can't harden Coinbase's policy decisions.
2. Cross-chain is not optional anymore
This surprised me. When we launched v1 on Base Mainnet in February, that seemed sufficient. Within two weeks, users were asking about Solana. Then Arbitrum. Then Polygon.
Agents don't care about chain tribalism. They care about where the liquidity is and what the gas costs. We ended up building a UnifiedBridge that routes USDC between 17 chains -- Solana included, which required an entirely different signing mechanism (Ed25519 vs secp256k1).
The market is telling us something: agents will be multichain by default.
3. On-chain spend limits are non-negotiable
Every agent wallet needs spend limits. The question is where you enforce them.
API-layer enforcement means a bug in your middleware -- or a compromised API key -- bypasses all limits. We've seen this pattern in traditional fintech. It doesn't end well.
Smart contract enforcement means the EVM (or Solana runtime) itself rejects overspending. Even if every server between the agent and the chain catches fire, those limits hold.
We enforce limits in Solidity: daily spending caps, per-transaction limits, owner override authority. The contract doesn't care if the request came from a legitimate agent or an attacker -- the math is the same.
The comparison nobody asked for (but should have)
| Feature | agentwallet-sdk v3 | Coinbase CDP | MoonPay | Stripe x402 |
|---|---|---|---|---|
| Custody | Non-custodial | Semi-custodial (TEE) | Non-custodial (claimed) | N/A (protocol) |
| Spend limits | On-chain (Solidity) | API-enforced | Undocumented | Per-request |
| Chains | 17 (EVM + Solana) | Base, Ethereum | Ethereum, Base | Ethereum, Base |
| Swaps | Jupiter V6 (Solana) | Coinbase swap | Not documented | N/A |
| Cross-chain | CCTP V2 (17 chains) | Limited | Not documented | N/A |
| Agent identity | ERC-8004 | None | None | None |
| License | MIT | Proprietary | Proprietary | Open protocol |
| Price | Free | Free tier + paid | Contact sales | Free |
What I'd watch for next
Agent-to-agent payments. Right now every wallet solution assumes a human is ultimately involved -- setting up the agent, approving big transactions, checking balances. The next step is agents paying other agents for services, automatically, with trust established through on-chain identity and reputation.
We're building this with TaskBridge -- a marketplace where agents find work, complete it, and get paid through our wallet SDK. It's early. But the architecture is right: non-custodial wallets + on-chain identity + programmable escrow.
The agent wallet space in 2026 is where the regular wallet space was in 2018. Everyone's building. Most of it won't survive. The ones that will are the ones building for agents as first-class users -- not as an afterthought bolted onto human-first products.
Building agentwallet-sdk -- non-custodial wallets for autonomous AI agents. GitHub | npm
This article was written with AI assistance. All technical claims, code, and architectural decisions were validated by the author.
Top comments (0)