About this episode
In this episode of Bullish on Bitcoin Show, Alexei is joined by Dominik Harz — co-founder and CTO of BOB, and the mastermind behind Pakt, BOB's newly launched guardrails product for AI trading agents. Dom brings a rare combination of backgrounds: a PhD in multi-agent systems and AI, years of Bitcoin bridging research, and hands-on experience building BOB Gateway into one of the most advanced Bitcoin trading engines in DeFi.
This episode tackles the most important emerging question in crypto: as AI agents increasingly manage financial decisions, who watches the agent? From the horror stories of agents gone rogue, to the fundamental limitations of LLMs as financial decision-makers, to a live demo of Pakt refusing bad trades on Hyperliquid mainnet in real time — this is the most technically grounded and candid conversation on agentic finance safety you'll find in the space.
In this episode we discuss:
- Dom's origin story: from a PhD in multi-agent systems at Imperial to building BOB with Alexei and coming full circle with Pakt
- Why AI agents are terrible at trading without guardrails — the auto-research experiment that blew up a portfolio through 20x leverage hill-climbing
- Why using AI to govern AI is non-deterministic and fails 10-20% of the time — and why that's unacceptable for financial systems
- How Pakt works: a cryptographic firewall between your agent and your portfolio, using zero-knowledge proofs to enforce hard trading rules
- Why ZK proofs make Pakt's guardrails deterministic, auditable, and mathematically unbreakable — even by the agent itself
- "- Agent proposes, proof decides" — why hallucination and API changes can never override a Pakt constraint
- Pakt's state proofs: the only system that can enforce account-level leverage limits, not just per-trade size limits
- Live demo: buying BTC long and ETH short on Hyperliquid mainnet, then watching Pakt refuse an oversized trade in real time
- ZK proof speed: 1.5 seconds per proof, up to 10 batched trades per proof — already near real-time for most trading use cases
- Privacy in Pakt: how ZK proofs can make your strategy invisible even to the infrastructure running it
- Proving trading performance without revealing trades: how Pakt's receipts could replace unverifiable "I made 200x" claims on Twitter
- McKinsey projects agents will mediate $3–5 trillion in global commerce by 2030 — and why Gartner expects 40% of businesses to decommission their agents by 2027 due to governance failures
- AI agent safety spending projected to hit $8 billion by 2028 — and why Pakt is positioned at the centre of that market
- The fully autonomous agentic finance future: why humans will eventually hand their finances entirely to agents — and why guardrails are the only thing that makes that safe

