Skip to main content
Worked, runnable recipes for the common agent architectures are being battle-tested now and land here shortly after launch. What is coming:

Single-agent swaps

A copy-paste system prompt that makes one Claude, OpenAI, or custom-loop agent swap-capable over the CLI: the commands, the confirmation rules, and what to do on each error code, for any agent that can run shell commands.

Shared swap agent

One agent holds the wallet and the API key and swaps on behalf of the rest of the system; every other agent requests swaps in its own domain language. One wallet, one audit trail, one place to enforce spend caps: the pattern for a treasury, a shared budget, or a product wallet.

Wallet per agent

A fleet of agents, each holding its own wallet and swapping for itself: isolated funds and isolated blast radius, one API key serving the whole fleet you own, delegation once per wallet per chain.
Until then, everything the recipes build on is already documented: the Agent skill covers driving the CLI unattended, the TypeScript SDK covers programmatic integrations including apps where every user brings a wallet, and Errors covers what an agent should do on each code.