Skip to content

Mind Agent Runtime

Architecture Review & Scorecard

本文是阶段性架构评审与成熟度记录,不定义 Architecture Truth。系统级职责与边界以 ARCHITECTURE_SYSTEM.md 为准,ProxyMind 客户端内部架构以 ARCHITECTURE.md 为准。

System Architecture · Runtime Architecture · Durability · Compute

Architecture Grade

A+

Overall Architecture Score

9.6 / 10

███████████████████▎  9.6 / 10

Architecture Stage

Core Architecture Mature · Verification & Production Hardening


01 · Executive Summary

Mind Agent Runtime 已经从传统的 Agent Framework 演进为一套具有明确 Runtime Authority、Durable Lifecycle、Compute Boundary 和 Recovery Model 的完整 Agent Runtime Platform。

当前系统由三个独立但协同的 Runtime Domain 组成:

┌─────────────────────────────────────────┐
│                  Mind                   │
│                                         │
│     Agent Intelligence & Execution      │
│                Runtime                  │
└───────────────────┬─────────────────────┘
                    │
             Intent / Command
                    │
                    ▼
┌─────────────────────────────────────────┐
│               AppServer                 │
│                                         │
│       Durable Lifecycle & Control       │
│                 Plane                   │
└───────────────────┬─────────────────────┘
                    │
           Compute Capability
                    │
                    ▼
┌─────────────────────────────────────────┐
│                 Fabric                  │
│                                         │
│          Cloud Compute Runtime          │
└─────────────────────────────────────────┘

三个系统最核心的职责划分已经稳定为:

Mind
owns Intent, Local Execution and Experience.

AppServer
owns Durable Lifecycle Truth.

Fabric
owns Cloud Compute Execution.

而贯穿整个体系的最高架构原则是:

One Fact, One Authoritative Owner.

System Architecture 已经明确区分三套 Runtime 的 Authority、Identity、Persistence、Recovery 和 Failure Boundary,而客户端架构进一步把这些原则落实到 domain / ports / application / harness / stores / adapters / frontends 的实际包边界中。


02 · Architecture Score

Architecture Dimension Score Grade Assessment
System Boundary 9.9 A+ Mind / AppServer / Fabric 边界高度清晰
State Authority 9.9 A+ One Fact / One Owner 已形成核心架构法律
Lifecycle Model 9.8 A+ Turn / Interrupt / Queue / Terminal 模型成熟
Protocol Model 9.8 A+ Submit / Observe、Command / Fact 边界稳定
Concurrency & Ownership 9.7 A+ Single Writer、Lease、Fencing 模型完整
Recovery Model 9.7 A+ Observation / Execution / Compute Recovery 已拆分
Effect Safety 9.7 A+ Unknown Effect → Reconciliation 模型成熟
Dependency Architecture 9.7 A+ 内部依赖方向稳定且具备自动化守卫
Compute Architecture 9.4 A Fabric 与 Agent Lifecycle 解耦正确
Presentation Architecture 9.5 A+ Presentation 与 Lifecycle 已彻底拆分
Observability Architecture 9.1 A 模型正确,跨 Runtime Trace 仍有增强空间
E2E Proof 9.0 A 架构已成熟,剩余主要是系统级证明

Composite Score

System Boundary          ███████████████████▊  9.9
State Authority          ███████████████████▊  9.9
Lifecycle Model          ███████████████████▋  9.8
Protocol Model           ███████████████████▋  9.8
Concurrency              ███████████████████▍  9.7
Recovery                 ███████████████████▍  9.7
Effect Safety            ███████████████████▍  9.7
Dependency Design        ███████████████████▍  9.7
Presentation             ███████████████████   9.5
Compute Runtime          ██████████████████▊   9.4
Observability            ██████████████████▏   9.1
E2E Proof                ██████████████████    9.0

Overall · 9.6 / 10


03 · Mind Runtime

9.7 / 10

Mature Agent Execution Runtime

Mind 已经超出了普通 Agent Loop 的范畴。

传统模型通常是:

Prompt
  ↓
Model
  ↓
Tool
  ↓
Answer

Mind 当前实际模型更接近:

User / Frontend
       │
       ▼
Immutable Command
       │
       ▼
Session Single Writer
       │
       ▼
Run Actor
       │
       ├── Model
       ├── Tool
       ├── Approval
       ├── Hook
       └── Effect
       │
       ▼
Protocol / Durable Runtime
       │
       ▼
Canonical Projection
       │
       ▼
Presentation

客户端内部已经形成明确的职责层:

agent/
├── protocol
├── domain
├── ports
├── application
├── harness
├── stores
├── capabilities
├── adapters
└── composition

并且规定 Domain、Application、Harness 和外围 Adapter 的依赖只能沿固定方向流动,包架构还有自动化审计。

Key Strengths

Single Writer
Stable Identity
Immutable Command
Explicit Ownership
Durable Evidence
Effect Reconciliation
Protocol Isolation
Presentation Projection

尤其成熟的是:

Local Run ≠ Remote Turn

以及:

Presentation ≠ Lifecycle

这两条直接消除了 Agent Runtime 中非常常见的“客户端状态冒充服务端事实”和“UI 状态反向控制执行生命周期”问题。


04 · AppServer

9.8 / 10

Durable Agent Control Plane

AppServer 是整套 Runtime 最强的可靠性基础。

它解决的问题已经不是:

“一次请求能不能返回结果?”

而是:

“一旦一次 Agent Execution 被接受,这个执行生命周期是否拥有唯一、持久、可恢复的事实。”

核心结构:

Session
   │
   ▼
Execution Gate
   │
   ▼
Durable Turn
   │
   ▼
Worker Lease / Fencing
   │
   ▼
Execution
   │
   ▼
Finalizing
   │
   ▼
Terminal Fact

System Architecture 已经明确把 Remote Turn、Session Gate、Durable Queue、event_seq、Worker fencing 和 Terminal Fact 的 Authority 归属 AppServer。

Architecture Highlights

Command ≠ Fact

Interrupt Accepted
        ≠
Turn Interrupted

Worker ≠ Authority

Worker
  may disappear

Turn Fact
  remains durable

Execution ≠ Finalization

Execution
  cancellable

Finalization
  durable

HTTP Connection ≠ Lifecycle

Disconnect
    ≠
Turn termination

这些不是实现技巧,而已经成为正式 Architecture Law。


05 · Fabric

9.3 / 10

Clean Cloud Compute Runtime

Fabric 最大的架构价值,是保持了 Compute Runtime 的纯度。

它不是:

Remote Agent Runtime

也不是:

Second AppServer

而是:

Compute Substrate

System Architecture 已经明确:

Fabric owns

Sandbox
Compute Run
Compute Resource
GPU / Model Execution

但不得取得:

Agent Session
Turn
Interrupt
Steer
Durable Queue
Agent Terminal

因此即使 Mind 或 AppServer 调用了 Fabric:

调用关系不会改变 Authority Ownership。

这是非常重要的长期可扩展性基础。

未来:

Fabric

Local Sandbox

Private GPU Cluster

Other Cloud Compute Runtime

都可以继续存在,而不会迫使 Agent Domain 重写。


06 · Authority Model

9.9 / 10

这是当前整个架构最强的一项。

                  AUTHORITY MODEL

                 ┌───────────────┐
                 │     Mind      │
                 │               │
                 │    Intent     │
                 │ Local Runtime │
                 │  Experience   │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │   AppServer   │
                 │               │
                 │ Durable Truth │
                 │   Lifecycle   │
                 └───────┬───────┘
                         │
                         ▼
                 ┌───────────────┐
                 │    Fabric     │
                 │               │
                 │    Compute    │
                 │   Execution   │
                 └───────────────┘

三个系统拥有不同种类的事实。

系统不存在:

GlobalRuntimeState
SharedTurnState
ClientTurnMirror
FabricTurn
UniversalRecoveryManager

这样的第二 Authority。

System Architecture 还进一步禁止通过修 Bug 的方式逐渐引入这些全局聚合模型。

这说明架构不仅定义了:

应该怎么设计

也定义了:

以后不允许怎么长。

这是架构成熟的重要标志。


07 · Identity Architecture

9.8 / 10

三个 Runtime 的 Identity 已经彻底分离:

Mind
────────────────
session_id
run_id
command_id
idempotency_key


AppServer
────────────────
cid
sid
turn_id
attempt
request_id
client_message_id
event_seq


Fabric
────────────────
sandbox_id
fabric_run_id
compute identity

核心原则:

Mind Run
   ≠
Remote Turn
   ≠
Fabric Compute Run

跨系统只通过明确 Adapter 建立映射,而不是依赖 ID 字符串偶然相等。

Identity Boundary 的成熟,直接支撑了:

Idempotency
Recovery
Replay
Fencing
Deduplication
Reconciliation

08 · Durable Lifecycle

9.8 / 10

一个标准在线 Turn 已经形成非常完整的生命周期:

User
 │
 ▼
Mind Command
 │
 ▼
Freeze Request
 │
 ▼
Submit
 │
 ▼
AppServer Durable Turn
 │
 ▼
Worker
 │
 ▼
Provider / Tool
 │
 ├──── Optional Compute ───► Fabric
 │
 ▼
Runtime Event
 │
 ▼
AppServer Event Plane
 │
 ▼
Mind Observe
 │
 ▼
Projection
 │
 ▼
User

最关键的分离是:

Submit
  ≠
Observe

新 Turn:

Submit
  ↓
Turn Created
  ↓
Observe

已有 Turn:

Attach / Replay
      ↓
Observe

Architectural Law

Existing Turn must never be resubmitted.

这条规则对解决网络响应丢失、Queue Start、Cold Recovery、Reconnect 场景中的 Duplicate Turn 极其重要。


09 · Recovery Architecture

9.7 / 10

Recovery 已经不再被抽象成一个模糊的“恢复机制”。

现在被明确分成:

Observation Recovery
────────────────────
Mind + AppServer

disconnect
   ↓
attach
   ↓
replay
   ↓
observe


Execution Recovery
──────────────────
AppServer

worker lost
   ↓
lease expires
   ↓
claim
   ↓
checkpoint
   ↓
continue / finalize


Compute Recovery
────────────────
Fabric

compute failure
   ↓
compute state
   ↓
retry / reconcile

三者拥有完全不同的 Owner。

Recovery Boundary follows Authority Boundary.

这一点对降低未来 Runtime Complexity 非常关键。


10 · Effect Safety

9.7 / 10

Agent Runtime 真正危险的不是模型回答错,而是:

外部副作用已经发生,但系统不知道它有没有发生。

当前架构已经明确:

Prepare
   ↓
Execute
   ↓
Commit

如果结果未知:

UNKNOWN
   ↓
Reconciliation

而不是:

UNKNOWN
   ↓
FAILED
   ↓
Retry

核心 Architecture Law:

Unknown Effect ≠ Failed Effect.

客户端内部同样把 Approval、Effect、Tool Call 和传输幂等身份拆开,不通过一个 request_id 冒充所有业务身份。


11 · Presentation Architecture

9.5 / 10

Mind 的 TUI 已经不是简单消费 Streaming Text。

当前展示链路具备独立 Projection:

Runtime Event
      ↓
CanonicalItemReducer
      ↓
TurnActivityProjector
      ↓
reduce_turn_surface()
      ↓
TuiTurnSurfaceCoordinator
      ↓
Physical Terminal

内部还明确区分:

Content
Presentation
Activity
Output Control

因此:

Thinking hidden
      ≠
Turn ended

Assistant visible
      ≠
Turn terminal

SSE closed
      ≠
Turn completed

客户端架构已经把 UI Projection 从 Runtime Truth 中彻底剥离。


12 · Complexity Assessment

Runtime Complexity

8.5 / 10

此分数表示系统本身复杂度,不代表质量高低。

Mind          7.8
AppServer     9.2
Fabric        6.7

System        8.5

复杂度主要来自:

Concurrency
Interrupt
Steer
Durable Queue
Replay
Cursor
Worker Lease
Fencing
Finalization
Effect Reconciliation
Cross-process Recovery
Presentation Coordination

但目前整体复杂度已经有一个非常重要的变化:

Before

Complexity
   ↓
hidden in race conditions


Current

Complexity
   ↓
explicitly modeled
   ↓
Authority
Identity
State
Lifecycle
Invariant

因此当前的复杂度主要属于:

Essential Complexity

而不是无意义的:

Accidental Complexity

Complexity Quality

9.3 / 10

系统虽然复杂,但复杂性已经基本做到 可命名、可定位、可解释、可测试


13 · Architecture Maturity

可以把 Agent 系统成熟度划成五级:

L1
Agent Script
│
│ Prompt + Tool
│
▼
L2
Agent Framework
│
│ Context + Session + Tool
│
▼
L3
Agent Runtime
│
│ Lifecycle + Execution + State
│
▼
L4
Durable Agent Runtime
│
│ Replay + Recovery + Idempotency
│
▼
L5
Agent Runtime Platform
│
│ Control Plane
│ Compute Plane
│ Observability
│ Governance
▼

当前 Mind:

                L5
                 │
                 │       ●
                 │    Mind Agent Runtime
                 │
─────────────────┼──────────────────
                L4

Current Maturity

L4.7 / L5

已经具备:

✓ Agent Runtime
✓ Durable Lifecycle
✓ Command Idempotency
✓ Replay
✓ Recovery
✓ Effect Reconciliation
✓ Durable Queue
✓ Control Plane
✓ Compute Plane
✓ Architecture Governance

距离完整 L5,剩余主要集中在:

Production Proof
Cross-Runtime Observability
Long-running Soak
Release Automation
Operational Experience

14 · Architecture Risk

现在最大的架构风险已经不是:

“设计方向是不是错的?”

而变成:

“实现能否长期保持这些 Architecture Laws 不被破坏?”

风险重心已经完成迁移:

Architecture Risk
      ↓↓↓

Implementation Drift
      ↓

Integration Risk
      ↓

Operational Risk
      ↑↑↑

因此下一阶段最应该保护的是:

Authority
Identity
Lifecycle
Invariant

而不是继续寻找新的 Runtime 抽象。


15 · Verification Gate

完整系统级验收应该继续围绕:

Normal Turn
Multi Turn
Long Streaming

Interrupt Thinking
Interrupt Assistant
Interrupt Tool
Interrupt Approval

Multiple Steers
Steer + Interrupt

Queue Add
Queue Start
Queue FIFO

Disconnect
Reconnect
Attach
Replay

Worker Failure
Finalizer Takeover

Fabric Success
Fabric Timeout
Fabric Failure

Terminal → Next Turn

核心指标:

Duplicate Turn       = 0
Duplicate Tool       = 0
Duplicate Terminal   = 0

Input Loss           = 0
FIFO Violation       = 0

False Cursor Gap     = 0
Cursor Corruption    = 0

Session Gate Stuck   = 0
Provider Replay      = 0

Unknown Effect Retry = 0

16 · Architecture Laws

整个 Mind Agent Runtime 可以最终压缩成这些不变量:

One Fact
    =
One Authoritative Owner


Command
    ≠
Fact


Submit
    ≠
Observe


Presentation
    ≠
Lifecycle


HTTP Ack
    ≠
Terminal


Connection
    ≠
Execution


Local Run
    ≠
Remote Turn


Remote Turn
    ≠
Compute Run


Worker
    ≠
Authority


Cache
    ≠
Truth


Unknown
    ≠
Missing


Unknown Effect
    ≠
Failed Effect


Compute
    ≠
Agent Lifecycle


Projection
    ≠
Source of Truth

这些原则已经同时贯穿系统级设计与 ProxyMind 内部实现约束。


17 · Final Score

╔════════════════════════════════════════════╗
║                                            ║
║           MIND AGENT RUNTIME               ║
║                                            ║
║        ARCHITECTURE SCORECARD              ║
║                                            ║
║              A+ · 9.6 / 10                 ║
║                                            ║
╚════════════════════════════════════════════╝
Area Status
System Architecture ✅ Mature
Runtime Architecture ✅ Mature
Authority Model ✅ Mature
Lifecycle Model ✅ Mature
Protocol Boundary ✅ Mature
Recovery Model ✅ Mature
Effect Model ✅ Mature
Compute Boundary ✅ Mature
Architecture Documentation ✅ Mature
E2E Proof ◐ Closing
Observability ◐ Hardening
Production Validation ◐ In Progress

Final Verdict

Mind Agent Runtime 的核心架构已经成熟。

现在形成的不是简单的三个仓库组合,而是三个拥有独立 Authority 的 Runtime:

Mind
Agent Intelligence & Execution Runtime

AppServer
Durable Lifecycle & Control Plane

Fabric
Cloud Compute Runtime

三者通过稳定 Identity、Command、Event 和 Capability Boundary 协作,而不共享一套模糊的 Global State。

因此下一阶段已经不应该继续以:

Architecture Redesign
Large Runtime Refactor
New Coordinator
New Global State

作为主要演进方向。

更合理的工程阶段是:

                Architecture
                    FROZEN
                       │
                       ▼
               ┌──────────────┐
               │ Verification │
               └──────┬───────┘
                      ▼
               ┌──────────────┐
               │ Stabilization│
               └──────┬───────┘
                      ▼
               ┌──────────────┐
               │ Observability│
               └──────┬───────┘
                      ▼
               ┌──────────────┐
               │  Production  │
               └──────────────┘

Architecture Signature

Mind thinks and orchestrates. AppServer makes execution durable. Fabric provides the compute substrate.

Final Rating

A+ · 9.6 / 10