- Rule-based rebalancing & yield rotations - Claim/compound + simple arb - Monitoring + alerts with hard limits - Research digests
Still hard on-chain: - Real strategy (not just execution) - Novel exploits & adversarial environments - Cross-chain reliability - Expressing nuanced risk prefs
What would make me give it more capital: - Provable bounds it *can’t* break - Full decision logs - Gradual autonomy + easy kill switch - Real track record under attacks - Skin in the game
Until then: boring stuff to agents, big decisions stay with humans.
AI agents can research and plan, but paying is still the hard part. Most systems force you to share a card or an open wallet — risky when the user is software.
Kite’s Agent Passport solves this. You approve one scoped session with a passkey: budget, per-tx limit, time window, and allowed destinations. The agent pays in stablecoins only inside those rules. Outside them, the tx fails. No private keys, no credit cards, full audit trail.
That’s the infrastructure the agent economy needs — machines that can act, with humans still in control.