Three-Layer Controllability

Core Question: Can I Trust AI

The biggest obstacle to letting AI help you isn't "can it do it," but "do I dare let it do it."

• "Is it doing something behind my back that I don't know about?"
• "What if it makes a mistake?"
• "If it messes up, can it be fixed?"

DesireCore answers these questions through the Three-Layer Controllability model, letting you enjoy AI efficiency while always maintaining initiative.

Three-Layer Controllability Model

          ┌───────────────┐
          │  L3: Reversible│  ← Can fix mistakes
          │  Reversible   │
          └───────┬───────┘
        ┌─────────┴─────────┐
        │    L2: Controllable│  ← Can decide what it does
        │    Controllable   │
        └─────────┬─────────┘
    ┌─────────────┴─────────────┐
    │       L1: Visible          │  ← Can see what it's doing
    │       Visible             │
    └───────────────────────────┘

L1: Visible — "I Can See What It's Doing"

Core Meaning: All agent behavior is completely transparent to you.

Specific Manifestations:

• Real-time status visible: What the companion is doing, how it's progressing — you can see anytime
• Decision process visible: Why the companion made this decision, what rules it followed, what memories it referenced
• Modification content visible: Every file modification by the companion is shown as diff (red for deletion, green for addition)
• Tool calls visible: Which tools the companion used, what parameters were input, what results were obtained

Analogy: Like an employee working with the door open — you can walk by anytime to see what they're doing.

L2: Controllable — "I Can Decide What It Can Do"

Core Meaning: You can intervene, confirm, or block agent behavior at any time.

Specific Manifestations:

• Permission grading: Different operations have different permission requirements
  • allow: Automatically allowed (low-risk operations)
  • ask: Ask every time (default mode, sensitive operations)
  • deny: Automatically denied (prohibited operations)
• Real-time interruption: You can pause or terminate executing tasks at any time
• Human Gate: High-risk steps automatically pause, waiting for your confirmation
• Configurable rules: "Allow it to read files, but ask me before writing files"

Analogy: Like giving employees different levels of access cards — some places they can freely enter, some places require you to swipe the card.

L3: Reversible — "If It Does Wrong, I Can Change It Back"

Core Meaning: Any operation can be undone or rolled back.

Specific Manifestations:

• Step-by-step rollback: Undo single modifications (Patch level)
• Turn-level rollback: Return to state before a certain conversation turn (Turn level)
• Session rollback: Return to state before the entire session began (Session level)
• Version snapshot: Can return to any historical version of the agent at any time

Analogy: Like a document's "undo" function, but covering all agent behavior — not just undoing text modifications, but undoing operations, undoing learning, undoing memories.

Why These Three Layers Are So Important

These three layers form a trust ladder:

Visible → Understand behavior → Establish preliminary trust
  ↓
Controllable → Set boundaries → Trust further enhanced
  ↓
Reversible → Eliminate worries → Complete trust
  ↓
Dare to delegate more tasks → Gain greater value

All three layers are indispensable:

• Only visible but not controllable? Can see but can't manage, equals forced旁观
• Only controllable but not reversible? Can manage but can't recover, mistake cost too high
• Only reversible but not visible? Can recover but can't see, don't know when to recover

Comparison with Other AI Products

Controllability Dimension	Traditional Chat AI	Low-code/RPA	DesireCore
Behavior Visibility	Only see output	See flowchart	See every step in real-time
Permission Control	None	Fixed flow	Flexible grading
Real-time Interruption	Can only stop generation	Can pause flow	Interrupt any operation anytime
Rollback Capability	None	Re-run flow	Step-by-step/turn-level/session-level rollback
Decision Transparency	Black box	See rules	See rules + AI decision basis

Controllability in Practice

In daily use, three-layer controllability manifests as:

1. Confirmation dialogs: When the companion wants to do sensitive operations, a confirmation box pops up telling you the operation content, risk level, and impact scope
2. Diff preview: Before the companion modifies files, it shows the diff first — you review item by item before applying
3. Receipt system: Every task has complete records, checkable anytime
4. Version history: Every change to the agent is traceable and rollback-capable

Next Steps

• Want to understand specific step types? Read Hardened/Flexible/Human Gate
• Want to understand task audit records? Read Receipt System