Most agent-accountability proposals hit the same wall: full cryptographic proof of every reasoning step an AI agent takes is too expensive to run in real time. A new architecture paper on distributed general-purpose agent networks proposes a workaround — split verification into three tiers instead of one all-or-nothing proof. The first tier runs at startup: the agent proves its loaded code and configuration match a registered commitment before it does anything else. The second tier runs during normal operation, where the agent keeps an append-only behavior log and periodically generates audit evidence tied to that registered configuration — batched and asynchronous, so it doesn't block real interactions. The third tier only kicks in when something goes wrong: a dispute or arbitration scenario can demand a stronger, more expensive proof for one specific interaction trace. The reasoning behind the split is pragmatic rather than ideological. The paper's authors are explicit that full zero-knowledge proof of every LLM reasoning step is currently too expensive for routine operation, so cryptographic evidence functions as an accountability substrate rather than a blanket requirement that all reasoning be proven online in real time. In effect, the system defers cost: cheap checks run constantly, expensive checks run only when accountability is actually contested. That's a meaningfully different design philosophy from identity-binding schemes that try to cryptographically pin every action to a verified operator up front. Instead of treating verification as a fixed cost paid on every transaction, this treats it as insurance — mostly dormant, activated on demand. The researchers built a prototype specifically to test whether this tiered approach holds up under real overhead, which is the right question: audit architectures that look elegant on paper often collapse once you measure the latency and compute they actually add to an agent pipeline. The unresolved part is what happens in the gap between tiers two and three — between a lightweight behavior log and a full dispute-grade proof. If the log itself can be manipulated or selectively omitted before a dispute triggers tier three, the whole chain's evidentiary value depends on how tamper-resistant that middle layer really is. The paper treats this as future work rather than a solved problem, which is a reasonable admission given how early this space still is. Still, the core insight — that accountability infrastructure needs a cost curve, not a single verification bar — is one more protocols in this space will likely have to reckon with as they move from whitepaper to deployment.