You've spent the rest of this academy learning to operate a wallet yourself. The AI agent course flips that: you authorise the wallet to act on your behalf, automatically, within rules you set, using a Large Language Model to read context and decide what to do.
That's a category of software that didn't exist five years ago, and it carries failure modes that crypto users never had to think about.
What "the agent" is, mechanically
The agent is two LLM calls (a planner and an executor) stitched together by deterministic Python code:
- Planner (Claude Opus) reads your prompt + your portfolio + the current allowance, then emits a structured JSON describing what to do
- A deterministic validator checks the plan against your config (cap, enabled use cases, address block-list)
- Executor (Claude Sonnet) receives the validated plan and dispatches it to one of five tools
Every phase is logged to an append-only ledger called agent_action_log and nightly Merkle-anchored on XRPL so a regulator (or you, with the signed audit bundle) can reconstruct exactly what the agent did and why.
Why metered allowance, not full delegation
A naive design would just be "AI agent with your seed phrase." That's broken: the LLM is the weakest link in any agent, and giving it unlimited authority means a single prompt-injection compromises everything.
Gopnik's design forces you to choose:
- Daily cap. EUR per UTC day. Hits zero → no more actions until tomorrow.
- Per-tx cap. Single largest action the agent can take. Hard ceiling separate from daily.
- Enabled use cases. Four boxes: x402 auto-pay, DCA scheduled buys, auto-bridge, portfolio Q&A. Disable any of them and the agent literally can't do that thing.
- Address block-list. Future iteration. For now: every destination is checksum-validated; suspicious patterns surface in the action log.
If you start with the default €5/day cap and never raise it, the worst-case "agent went rogue" outcome is €5 of misallocation. The audit bundle tells you which prompt caused it and what plan the planner emitted.
The four use cases shipped in iter-E
x402 auto-pay. The agent pays for HTTP 402 endpoints automatically. Useful for paid APIs, on-demand compute, data feeds. The next lesson covers x402 in depth.
DCA scheduled buys. "Buy €25 of XRP every Monday from my USDC balance." The agent persists a schedule; a background worker fires individual buys on cadence.
Auto-bridge rules. "When my Base USDC exceeds €1,000, bridge €500 to Arbitrum via CCTP." The agent persists the rule; a worker evaluates rules every 5 minutes.
Conversational portfolio Q&A. Read-only. "Why is my Polygon balance lower than last week?" The agent answers from a structured snapshot; never spends money in QA mode.
What you commit to with this certification
- Each agent action is logged, including the planner's input and the executor's output
- Daily and per-tx caps are your job to set sanely
- Disabling the agent stops new actions but doesn't reverse in-flight ones
- A signed JSON bundle of your day's actions is downloadable any time
The next lesson covers HTTP 402 — the protocol that turns "pay for an API" into a primitive.