Psy as Agent Substrate 作为智能体基质的 Psy

System Decomposition系统分解

Six primitives compose the current Psy stack. For each, I'll mark what's shipped and what's missing if we want this to coordinate millions of autonomous agents — not just hold their wallet keys. 当前 Psy 技术栈由六大原语组成。对每一项，我将标注「已交付」与「若想协调数百万自主智能体（而非仅托管其钱包密钥）所缺失的」。

The deepest gap: Psy treats agents as holders of accounts. To coordinate millions of them, agents must become first-class principals — entities whose memory, capabilities, history, reputation live as native chain state, not as JSON in some sidecar database. 最深层的缺口：Psy 把智能体当作账户的持有者。要协调数百万智能体，它们必须成为一等公民——其记忆、能力、历史、声誉作为原生链状态存在，而非寄居于某个旁路数据库的 JSON。

Application Forecast — Twenty High-Impact Bets应用预测——二十项高影响力押注

Sorted into four categories. Each is built directly on primitives the public stack already ships (or that the v2 architecture in §5 would add). None are vapor; each names the repos it depends on. 分为四类。每一项均直接构建于公开栈现有原语（或第 5 章 v2 架构将引入的原语）之上。无一为概念炒作；每项均标注所依赖的仓库。

Emergent Behaviors at Million-Agent Scale百万智能体规模下的涌现行为

When you have five agents, you have software. When you have five million, you have an ecosystem with its own dynamics — most of which were never designed in. Some are gifts; some will eat the system if not anticipated. 当你有五个智能体，那是软件。当你有五百万，那是一个拥有自身动力学的生态——其中绝大多数从未被设计在内。有些是礼物；若不预先识别，有些会吞噬整个系统。

Missing Layers — What "Living Network" Requires缺失的层——「活体网络」所需

The current stack ships data, state, computation, consensus. To evolve into something agents can live in, eight more layers are needed. Four are load-bearing; four are quality-of-life. 当前栈交付了数据、状态、计算、共识。要演化成智能体可以居住其中之物，还需八层。四层承重，四层关乎生活质量。

The four most load-bearing for the agent thesis: Memory · Reputation · Intent · Capability. Without all four, agents remain wallets-with-extra-steps. 对智能体论题最承重的四层：记忆 · 声誉 · 意图 · 能力。四者缺一，智能体就仍然只是「带额外步骤的钱包」。

V2 Architecture — A Paradigm ShiftV2 架构——范式转变

The shift, in one sentence: from a transaction-oriented chain that agents use → to an intent-and-context-oriented substrate that agents are made of. 用一句话概括转变：从智能体所「使用」的、面向交易的链 → 转向智能体「身体所构成」的、面向意图与上下文的基质。

┌────────────────────────────────────────────────────────────┐
│                INTENT NETWORK (new top layer)              │
  Agents post WHAT; solvers compete on HOW; settlement on
  outcome. Cross-chain by default. CoW + ZK + agent-first.
├────────────────────────────────────────────────────────────┤
│                SEMANTIC MEMORY FABRIC                      │
  ZK-verifiable indexed knowledge. Micropriced retrieval.
  RAG-as-network-primitive. Memory becomes consensus state.
├────────────────────────────────────────────────────────────┤
│  REPUTATION VECTORS      │     CAPABILITY GRANTS           │
  Multi-dim, decay-aware,    Bounded ZK permissions with
  cross-chain portable       recursive sub-grant trees
├────────────────────────────────────────────────────────────┤
│              AGENT IDENTITY GRAPH                          │
  Composable: agent → operator → KYC root. Selective
  disclosure at any depth. Identity is a tree, not a key.
├────────────────────────────────────────────────────────────┤
│         NEGOTIATION + LIVENESS PRIMITIVES                  │
  Off-chain commit-reveal handshakes · Native scheduler
  Deadman switches · Auction protocols · Cycle detection
├────────────────────────────────────────────────────────────┤
│           SPECIALIZED REALM HIERARCHY (v2 sharding)        │
  Compute │ Memory │ Intent │ Oracle │ Settlement realms,
  each with proof system tuned to workload. Heterogeneous.
├────────────────────────────────────────────────────────────┤
│            PROOF-OF-RESULT CONSENSUS                       │
  Validators stake on outcome correctness, not block
  sequence. Verify the meaning, not the transcript.
├────────────────────────────────────────────────────────────┤
│       RECURSIVE PROOF FABRIC (Plonky2.5 → Plonky3)        │
  Same proof system spans all layers. One global verifier
  proves the entire stack as one state transition.
└────────────────────────────────────────────────────────────┘

Five Concrete Decisions That Actually Matter五个真正重要的具体决策

1. Heterogeneous realms.异构分片域。 Instead of 128 identical shards, ship ~6 realm types, each with a proof system tuned for its workload. Plonky2.5 makes this possible — the recursive layer collapses heterogeneity at verification time.不再使用 128 个等同分片，而是约 6 类分片，各自配备针对其工作负载调优的证明系统。Plonky2.5 让这成为可能——递归层在验证时消解异构性。
2. Memory as base state.记忆作为基础状态。 Add a third state-tree alongside balance and contract state: a Merkle-DAG of semantic embeddings with provable retrieval. Single biggest expansion of what "chain state" means.在余额与合约状态之外增加第三棵状态树：附可证明检索的语义嵌入 Merkle-DAG。「链状态」含义最大的一次扩展。
3. Capability tokens, not just balance.能力代币，而不仅是余额。 Every transaction needs a capability grant, not just a signature. Wallets become root capability holders; agents inherit bounded subsets.每笔交易都需要一个能力授权，而不仅仅是签名。钱包成为根能力持有者；智能体继承受限子集。
4. Outcome-settlement, not transaction-settlement.按结果结算，而非按交易结算。 Don't settle "agent A sent $X to B." Settle "A's intent buy_under($Y) was fulfilled by solver C, verified outcome F."不要结算「A 向 B 转 $X」。结算「A 的意图『以低于 $Y 买入』被求解器 C 实现，验证结果为 F」。
5. Bond-cycle consensus rules.担保环路共识规则。 The chain refuses to confirm bond chains with cycles or excessive depth. Recursive guarantees enforced at the protocol layer prevent LUNA-style cascades.链拒绝确认含环路或深度过大的担保链。在协议层强制执行的递归保证可阻止 LUNA 式级联崩溃。

The Bold Insight大胆洞见