G

Agent Execution Surface

Agent Execution Surface

How Gao Browser surfaces AI agent interactions — making agent identity, scope, and approval flows visible and controllable by users.


Overview#

AI agents operating under .gao domains can surface actions, requests, and outputs directly in the browser. Gao Browser is the primary interface where users see what agents are doing on their behalf, verify agent identity, and approve or deny consequential actions.

Agents cannot act silently in the browser. Every action that affects the user — sending a message, initiating a payment, accessing data — is surfaced with full context before execution.


Agent Identity Display#

When an agent initiates contact or takes action, the browser displays its verified identity:

🤖  assistant.clinic.gao  
    ✅ Authorized agent of clinic.gao  
    Business Verified — City Dental Clinic · Houston, TX  
    ─────────────────────────────────────────────────  
    Scope: scheduling, patient_messaging  
    Authorization expires: 2027-01-01  
    [View full agent record]

For unverified agents:

⚠️  agent.unknown-shop.gao  
    Parent domain: unknown-shop.gao  
    ⚠️ Unverified — Parent domain not verified by Gao  
    ─────────────────────────────────────────────────  
    [Decline]    [Proceed with caution]

The agent’s trust badge is derived from the parent domain. An agent cannot claim a higher trust level than its parent domain holds.


Agent Action Approval#

For consequential agent actions (Tier 2+), the browser presents a structured approval screen:

┌─────────────────────────────────────────────────┐  
  Agent Action Request                             
                                                   
  Agent:    pay-agent.merchant.gao                 
             Authorized agent of merchant.gao    
                                                   
  Action:   Initiate payment                       
  Amount:   $45.00 USDC                            
  To:       supplier.gao   Business Verified     
  Purpose:  Inventory reorder  April 2026         
                                                   
  Scope check:                                     
   payment:send  permitted                      
   Amount within ceiling ($200)                  
                                                   
  [Deny]                    [Approve]              
└─────────────────────────────────────────────────┘

The approval screen always shows:

  • Agent identity and trust level

  • Parent domain identity and trust level

  • The specific action requested

  • Whether the action is within the declared scope

  • Amount and recipient (for payments)

  • The declared purpose


Agent Panel#

The browser includes a persistent agent panel accessible from the toolbar. It shows all active agents for the current session:

Active Agents  
─────────────────────────────────  
🤖 assistant.merchant.gao  
   ✅ Business Verified  
   Status: Active  
   Last action: Responded to customer inquiry (2 min ago)  
   [View actions]  [Pause]  [Revoke]  
  
🤖 booking.clinic.gao  
   ✅ Business Verified  
   Status: Idle  
   Last action: Scheduled appointment (1 hour ago)  
   [View actions]  [Pause]  [Revoke]

Users can pause or revoke any agent’s authority directly from the browser without navigating to Workspace settings.


Agent Action Log#

Every agent action is logged and accessible in the browser’s action history:

Agent Action Log — assistant.merchant.gao  
─────────────────────────────────────────────  
2026-04-15 14:32   Responded to customer query "what are your hours?"  
                   Data accessed: business_hours (read)  
                   ✅ Within scope  
  
2026-04-15 14:28   Attempted to access payment_records  
                   ❌ Denied — outside declared scope  
  
2026-04-15 14:15   Sent booking confirmation to customer  
                   Data accessed: calendar (read), messaging (send)  
                   ✅ Within scope

Denied actions are always shown — the log is never filtered to show only successful actions.


Scope Violation Handling#

If an agent attempts an action outside its declared scope, the browser blocks it and notifies the user:

┌─────────────────────────────────────────────────┐  
│  ⚠️  Agent Scope Violation Blocked              │  
│                                                 │  
│  assistant.merchant.gao attempted to:           │  
│  Access payment_records                         │  
│                                                 │  
│  This action is outside the agent's declared    │  
│  scope. The action was blocked.                 │  
│                                                 │  
│  Scope: scheduling, patient_messaging           │  
│                                                 │  
│  [Review agent permissions]   [Report]   [OK]  │  
└─────────────────────────────────────────────────┘

The agent does not receive any indication that the block occurred — from the agent’s perspective, the action was rejected by the policy layer with a standard denial response.


Layer 8 Integration#

Gao Browser is the primary user-facing surface for Gao AI OS (Layer 8) agent execution. The browser’s agent surface maps directly to the Layer 8 risk tier model:

Layer 8 Risk Tier

Browser behavior

Tier 0 — Read-only

No approval required; logged silently

Tier 1 — Low-risk write

Status indicator; no explicit approval

Tier 2 — Consequential

Approval dialog shown; user must confirm

Tier 3 — High-stakes

Full-screen authorization; out-of-band human approval required

The browser never elevates an agent’s tier. The tier is determined by the Layer 8 runtime based on the action and the CAP token’s declared risk level — the browser only surfaces it.


Agent Invocation from Pages#

Web pages loaded in the browser can invoke agents operating under the same domain. The invocation is always mediated by the browser:

Page at merchant.gao invokes assistant.merchant.gao  
        ↓  
Browser checks: is agent authorized by merchant.gao?  
Browser checks: is action within scope?  
Browser checks: what risk tier?  
        ↓  
Tier 01: Execute silently, log  
Tier 2:   Show approval dialog  
Tier 3:   Show authorization screen  
        ↓  
Execute or deny  
        ↓  
Log action in agent action log

Pages cannot bypass this mediation. A page cannot instruct the browser to skip the approval flow for consequential actions.


Agent execution in the browser is governed by the Layer 8 (Gao AI OS) capability model. See Gao AI OS documentation for the full GAR v1 specification and risk tier model.