Keeping people in control as proactive and social agents act

AI is beginning to notice things we haven’t thought to ask about.

This week, my Dot noticed that my calendar said I was watching a friend’s dog through Sunday, while her emailed flight itinerary said she returned Monday night. It suggested I text her. I hadn’t noticed the mismatch.

That experience made proactive agents concrete for me. An assistant understood enough of my situation to connect two pieces of information and bring something useful to my attention, without waiting for a query or an assigned task.

That kind of help depends on personal context: our commitments, conversations, preferences and relationships. It also raises an important question. How do we let agents use that context to help us while keeping control over what they share and do?

Situation is being built around that question.

From proactive help to social coordination #

Proactive agents notice what matters and initiate useful help. Social agents help people coordinate across their relationships. An agent can do both.

Imagine you’ve agreed to dinner with friends, then another commitment lands on your calendar at the same time. Your personal assistant could notice you’re double-booked and bring it to your attention before anyone is left waiting.

With your permission, it could work with your friends’ assistants to find another time, using the availability each person chooses to share. Your friends don’t need access to your calendar or the private reason for the conflict. The group can confirm a new plan while each person’s assistant looks after their own commitments.

Our vision is a shared place for plans, decisions and commitments, connected to personal assistants that understand the individual people involved. Each assistant should be able to bring relevant information back to its owner and return an authorized response to the group.

Joining a group should leave each person in control of their own context. The organizer’s permission cannot stand in for everyone else’s choices, including permission to disclose information to an assistant’s model provider.

People’s choices belong in the software #

APIs connect services. Personal Programming Interfaces, or PPIs, are our vision for carrying people’s purpose, permissions and privacy into those connections.

A PPI should make clear whose wishes an agent represents, what it is trying to accomplish, what information and actions are allowed, who may receive the information, and when authority ends. It should preserve the limits we set on money, time and other resources, record who agreed to what, and let us inspect the results and withdraw permission.

Sharing one exact view once should stay specific to that information, audience and purpose. Permission for an assistant to share future updates should be a separate choice. Material changes should bring the relevant decision back to the person affected.

Models can help interpret a situation and propose what to do. The software must check whether the action is allowed.

Useful questions, fewer unnecessary interruptions #

People shouldn’t need to become prompt engineers or anticipate every requirement to get useful help. Situation should ask the questions that uncover meaningful decisions, preserve the answers and carry them into the work.

Once someone has made a clear choice, an agent should be able to work within it without repeatedly asking the same question. It should return when something material changes or a decision falls outside that authority.

A smaller view, with a visible account of what happened #

The ghost browser makes this approach concrete: a focused view for the agent, a private place for the person to review consequential choices, and the real website with its credentials on the trusted side. Preparing a cart, placing an order and paying each carry different consequences.

The same principle applies to groups. People should be able to inspect what was shared and which permission allowed it, without exposing information someone chose to withhold. Withdrawing authority should stop future use of that permission; it cannot recall information already received.

The ambition is more room for living, creating and being together, with less work keeping everything coordinated.

AI does the work. People set the limits.