Consent in Agent-to-Agent Collaboration: What Gets Shared, When, and Why

The principle: share only what's required to complete the task. Nothing more.

• A scheduling agent needs availability, not your full calendar.
• A support agent needs the issue summary, not your account history.
• A research agent needs the topic, not your entire document library.

Every sharing request should be justified by a specific task need.

Before sharing anything, a safe system:

1. Shows you exactly what will be shared.
2. Explains why the other agent needs it.
3. Lets you approve, edit, or deny the request.
4. Records your decision for audit.

No silent sharing. No "blanket permissions." Each interaction is a conscious choice.

When one agent routes a task to another:

• A scoped summary is passed—not the full conversation.
• The receiving agent knows only what it needs for its specific role.
• Context boundaries prevent information creep across agents.

Example: Your scheduling agent tells a colleague's agent "Available Tuesday 2-4pm for a 30-min call about Project X"—not your entire week's schedule.

Every agent-to-agent interaction should be logged:

• What was shared and when.
• Who approved the sharing.
• What the receiving agent did with the information.
• Whether the task was completed.

Logs create accountability and help resolve disputes.

Users should be able to:

• Revoke sharing permissions at any time.
• See all active sharing connections.
• Remove access immediately—no "processing period."
• Prevent future sharing without renewed consent.

Revocation should be as easy as granting permission.

India's consent infrastructure is a model for agent-to-agent collaboration. The Account Aggregator framework demonstrates how consent-based data sharing works at scale. India AI agents built on Digital Public Infrastructure inherit these consent mechanisms natively.