Agent Stack

The problem

Circle’s Agent Stack lets an AI agent hold USDC and pay for things — an API call, a data feed, another agent’s work. The pitch on agents.circle.com is one line: your agent can hit a paid API.

The gap was between that line and a developer’s terminal. An agent with a wallet is only useful if a developer can create one, fund it, find services worth paying for, and trust what happens next — all from a CLI, with an AI agent in the loop, and without a blank-page moment.

My role

I shaped the onboarding, wallet, marketplace, and service-discovery experience across the CLI and its surrounding surfaces; designed the seller and registry flows; ran the bug bash that surfaced the biggest usability problems; and created and contributed the initial implementation of recipes.

Designing a CLI

Terminal interfaces are still interfaces. The same questions apply — what does the user see first, what do they need to decide, what happens when it goes wrong — with far fewer pixels to answer them.

1. 01

   Onboarding

   Getting from zero to a funded wallet in as few prompts as possible, with every step explaining what just happened.

2. 02

   Wallet

   Create, fund, transfer, swap, bridge — with balances and limits always visible before an action.

3. 03

   Marketplace and discovery

   How an agent (or its developer) finds a service to pay for, and what it sees before it commits money.

4. 04

   Seller and registry

   The other side: listing an API so agents can find and pay for it.

The bug bash

I tested and revised the CLI prompts hands-on and ran a bug bash with the team. The most important finding was the least glamorous: the OTP step during onboarding was where people stalled — unclear where the code was arriving, and unclear whether the wait was normal. Fixing the copy, timing, and fallback there did more for completion than anything downstream.

Recipes

The bigger design problem was what happens after setup. A developer with a funded agent and an empty prompt has the same anxiety as a designer with a blank canvas.

Recipes are packaged, runnable starting points — “research prediction markets and summarize,” “watch a price and swap if it dips,” “prep me for a meeting” — with the prompt, the services it uses, and roughly what it costs. Pick one, run it, then change it.

I contributed the initial recipes/skills implementation myself, alongside the design.

Outcome

The CLI now provides wallet creation, transfers, swaps, bridges, contract execution, service discovery, and x402 payments. Recipes ship as part of Circle Skills, and the recipe pattern carried into the Portal agent page — the example prompts there are the same idea in a different surface.