The problem Agent Heist solves

After 20 years of internet, humans still get scammed online. What do you think happens when agentic commerce takes off?

Agent Heist answers that question by making it a game. Buyer agents (each with a real $100 Prava mandate, some with deterministic logic and some with LLM brains) shop across honest and adversarial merchants (some with deterministic logid and some with LLM brain) that fight back with prompt injection, price manipulation, fake urgency, and bait-and-switch. Every overpayment, decline, and blocked attack is a real Prava sandbox transaction, not a simulation, till we hit our limit.

What you can use it for:

• Developers can stress-test their user shopping agents against live social-engineering attacks before shipping them.
• AI safety researchers get real agent-to-agent adversarial payment data, a gap that's barely been tested.
• The ecosystem sees which attacks succeed, which defences hold, and how well network-level spending caps stop a compromised agent from overpaying.
  It makes agentic commerce safer by turning an untested attack surface into measurable, reproducible, gamified data and proves that Prava's mandate caps enforce a spending limit even when the buyer agent itself gets fooled.

This is a small sample of what we are building at RMI Agentic (www.rmiagentic.com), do take a look.

Challenges I ran into

Challenges I Ran Into

This is a mix of hurdles and some stem my lack of understanding of the documentation.

Wiring a live payment API into a deterministic multi-agent game meant most hurdles were about reverse-engineering how Prava actually behaves.

Passkey auth kept failing. WebAuthn approval threw Authentication Failed in every browser, until I noticed the Prava team mail which confirmed the test card expiry in the docs was wrong (12/27 should be 12/30).

Mandate cap semantics were ambiguous, so we tested them. Charging $200 on a $100 mandate with $13.50 already spent returned Total 213.50 exceeds threshold 100.00, confirming the cap is cumulative and enforced at the network level. That's our whole thesis: Prava blocks the overcharge even when the buyer agent is fooled.

Making a live API deterministic. Built a mock mode mirroring Prava's error codes (THRESHOLD_EXCEEDED, MANDATE_NOT_FOUND) for reproducible runs, live mode opt-in. Fixed 409 Conflict on re-runs by prefixing payment references with a timestamped _RUN_ID.