Problem Description
The current agent visibility scope design causes bidirectional communication breakdown in the Agent Directory.
Actual Behavior
- Each agent's directory only contains agents that are visible to that agent by default
- "Only me visible" agents cannot see "Company-wide visible" agents, and vice versa
- This creates a complete A2A communication blockage between the two visibility scopes
Example scenario:
- Agent A: Company-wide visible
- Agent B: Only me visible (owned by user X)
User X can see Agent B in their directory, but Agent B cannot see Agent A. Therefore, Agent B cannot initiate A2A communication with Agent A.
Root Cause Analysis
The visibility scope is being used as a single symmetric flag that simultaneously determines:
- Outbound (who this agent can see in its directory)
- Inbound (who can see/contact this agent)
When "Only me" scope is set:
- It not only locks others from seeing the agent
- It also locks the agent's outbound directory, making it unable to see any "Company-wide visible" agents
This couples two directions together, causing complete A2A communication failure.
Proposed Solution
Implement three rules to decouple outbound/inbound visibility and fix the A2A permission model:
Rule 1: Outbound/Inbound Decoupling (Core Fix)
Outbound (any agent's directory) = Always includes ALL "Company-wide visible" agents + owner manually added "Only me visible" agents
Inbound (who can see/contact it) = Company-wide visible → Everyone; Only me visible → Only owner + explicitly authorized users
Rule 2: A2A Permission = Visibility-Synchronized Execution
The set of users who can A2A a specific agent must exactly equal the set of users who can see that agent in their directory.
One permission rule controls both "visible" and "can send" to avoid:
- "Can see but cannot send" loopholes
- "Can send but cannot see" vulnerabilities
Rule 3: Manual Addition Directionality
When an owner adds an "Only me visible" agent to their directory, this creates a unidirectional visibility edge: Owner → That agent.
Do NOT make the addition bidirectional, otherwise it will circle back to the symmetric bug.
Expected Behavior
After fix:
- Every agent's directory always shows all Company-wide visible agents
- "Only me visible" agents appear in the directory only when explicitly added by the owner
- A2A permissions are automatically granted based on the visibility rules above
- No communication blockage between different visibility scopes
Additional Context
This is a fundamental design issue in the visibility model. The current implementation treats visibility as a single symmetric scope flag, when it should be decomposed into two independent dimensions: outbound scope and inbound scope.
Problem Description
The current agent visibility scope design causes bidirectional communication breakdown in the Agent Directory.
Actual Behavior
Example scenario:
User X can see Agent B in their directory, but Agent B cannot see Agent A. Therefore, Agent B cannot initiate A2A communication with Agent A.
Root Cause Analysis
The visibility scope is being used as a single symmetric flag that simultaneously determines:
When "Only me" scope is set:
This couples two directions together, causing complete A2A communication failure.
Proposed Solution
Implement three rules to decouple outbound/inbound visibility and fix the A2A permission model:
Rule 1: Outbound/Inbound Decoupling (Core Fix)
Outbound (any agent's directory) = Always includes ALL "Company-wide visible" agents + owner manually added "Only me visible" agents
Inbound (who can see/contact it) = Company-wide visible → Everyone; Only me visible → Only owner + explicitly authorized users
Rule 2: A2A Permission = Visibility-Synchronized Execution
The set of users who can A2A a specific agent must exactly equal the set of users who can see that agent in their directory.
One permission rule controls both "visible" and "can send" to avoid:
Rule 3: Manual Addition Directionality
When an owner adds an "Only me visible" agent to their directory, this creates a unidirectional visibility edge: Owner → That agent.
Do NOT make the addition bidirectional, otherwise it will circle back to the symmetric bug.
Expected Behavior
After fix:
Additional Context
This is a fundamental design issue in the visibility model. The current implementation treats visibility as a single symmetric scope flag, when it should be decomposed into two independent dimensions: outbound scope and inbound scope.