Skip to content
Product Updates

Agent Wallet on Arc Testnet — first end-to-end flows are working

A testnet progress note from the Agent Wallet team. On Arc — a stablecoin-native, USDC-as-gas blockchain — the agent can already query balances, send by raw address or saved alias, register contacts, and walk a user through a swap with explicit confirmation. Still pre-release, still iterating.

369wallet team4 min read
369wallet Agent Wallet testnet update — first end-to-end flows working on Arc Testnet

It has been a couple of weeks since we opened early access to the Agent layer. The team has been heads-down on testnet — wiring tools, tightening the conversation loop, hardening permissions. This post is a short progress note: what the agent can already do end-to-end, captured directly from the build we're testing on Arc Testnet.

Three flows below. None of them are demos in the bad sense — they are screenshots from the build we are dogfighting daily.

Why we're testing on Arc

The screenshots use Arc Testnet. Arc is a stablecoin-native EVM-compatible chain — its key difference is that USDC itself is the native gas token, so transfers, swaps, and contract calls all settle without juggling a separate gas asset. For an agent that mostly moves stablecoins on behalf of a user, that removes an entire failure mode. It's a good first chain to harden the agent loop on; the same flows port to the other 30+ chains the Agent layer targets at GA.

Flow 1 — Balance, send, register

Three primitives in one short conversation. The user asks the agent for their USDC balance, then sends a small amount to a raw address, then asks the agent to save that address as a named contact. The agent returns structured tool-result cards rather than just text — a balance card, a "SENT" card with the on-chain tx hash, a contact-registry confirmation.

Agent chat on Arc Testnet: user asks for USDC balance, sends 1 USDC to a raw address, and registers that address as a contact named Alice. Each step returns a structured result card.

Three things matter in this exchange: the agent parses intent from plain language (no command syntax), it shows structured results inline (the SENT card with tx hash, not just a string), and state persists across turns (the same conversation later treats "Alice" as a known recipient).

Flow 2 — Send by name, ask "what's special about this chain?"

Because the address book persists, the next time the user wants to send, they don't have to remember a 42-character hex string. "Send 0.1 usdc to alice" resolves through the agent's contact registry and settles on-chain — the same SENT card pattern, but the intent is human-shaped.

In the same window, the user asks "What's special about Arc?" — and the agent answers in plain English (USDC as native gas, EVM-compatible, payments-optimized, currently testnet-focused). The point of including this isn't the trivia answer. It's that the agent's domain knowledge and its action surface are the same agent. You don't switch tools between asking and doing.

Agent chat on Arc Testnet: user sends 0.1 USDC to 'alice' by alias and then asks the agent about Arc's characteristics. The agent confirms the send with a tx hash card and explains Arc as a stablecoin-native EVM chain in plain language.

Flow 3 — Swap with explicit confirmation

Transfers are reversible only through goodwill; swaps are not. So the agent layer enforces an explicit confirmation step for value-transforming actions. When the user asks the agent to swap 0.1 USDC into EURC, the result is not a transaction — it's a review modal: the protocol it would route through (Circle Stablecoin Kits), the chain (Arc Testnet), the expected receive amount, the max slippage, and a clear Cancel / Confirm swap pair.

Agent chat on Arc Testnet showing the Review swap modal for a USDC → EURC swap: 0.1 USDC, Circle Stablecoin Kits protocol, Arc Testnet chain, ≈0.0982 EURC receive amount, 3% max slippage, and explicit Cancel / Confirm swap buttons. An earlier swap attempt was cancelled, which the conversation log shows.

This pattern is the design we are settling on for anything that materially changes the user's portfolio: the agent proposes, the user confirms, the chain settles. The earlier "Cancelled" entry in the same thread is intentional — we want it to be just as easy to back out as to commit.

What we are seeing converge

  • Address book as a first-class agent capability. Contacts are not a UI accessory — they materially shrink the surface area for catastrophic typos.
  • Structured tool results over chat strings. A SENT card with the explorer link is denser, more trustworthy, and less hallucinatable than a one-line "ok done" message.
  • Cancel parity for swaps. Every value-transforming action ships with a cancel that feels as native as the confirm.
  • Domain knowledge alongside actions. Same agent answers "what is this chain" and "send 0.1 USDC there." We do not want a second app for explanation.

What's not yet in

Some specific pieces are still on the testnet to-do list, in honest order:

  • Cross-chain routing inside agent intents (right now flows live within a single chain per turn).
  • Session-grant UI for scoped permissions — the SDK-side primitive is in; the user-facing approve/revoke surface is being designed.
  • x402 payment flows triggered by the agent for paid API calls.
  • Staking and prediction-market actions through the agent surface (the underlying integrations already work for human users).

Status

Testnet. Early access remains gated through the signup on the /en/agent page. We will keep posting these small progress notes as the loop tightens — and a longer write-up when the GA build is close.

"The screenshot test is the only test that matters. If you can't show the flow happening, you don't have the flow." — 369wallet team

— The 369wallet team