Quillow

10 Nov 2025

Personal AI agents that are proactive

We are an engineering-led consumer AI team building the next generation of personal AI agents that are truly proactive.

The best people have agency. They don't wait for instructions from their employer, client or friend. They anticipate what is required and just get it done. When you tell the best people what to do, they have already done it.

The same is true with AI. The future of personal AI agents can't be just one prompt after another.

In the future, everyone will have a personal AI that will know them extremely well and be able to help them proactively.

We want to help make this a reality at Quillow.

Agents need to understand time to be truly proactive

To deliver on this, certain critical (software) engineering challenges need to be solved.

At its core, the main technical problem we are attempting to solve is to get AI agents to understand time well. Current systems act on either direct prompts, specified workflows or agentic loops with a goal, but they do not really understand time.

To be proactive, the agent must be running continuously, have agency, deeply know the user and know how to reach the user at the right time all while maintaining privacy.

This gives us the five pillars that we are working on:

Memory

How do we best index information in a way that helps the AI agent to retrieve the right context at the right time? When something happened is as important as what happened.

Agency

How do we efficiently provide tools such as computer-use so that the agent can take autonomous action with guardrails?

Communication

How can we allow the agent to reach the user proactively? e.g. message them on Whatsapp, or SMS, or other apps.

Privacy

How can we ensure privacy? We certainly take this seriously, which is why we have a subscription based business model and have robust privacy policies that protect user's data. How can we achieve this techncially without compromising on the quality of the AI agents?

Continuous inference

How do we trigger the agent via a system clock (a.k.a cron jobs) in the most effective manner? We have leaned on this the most and have built robust infrastructure to support this.

Continous, useful inference for all

Ultimately, our mission is to provide personal AI that runs continously for all.

We believe there needs to be a model-agnostic service that delivers this.

Edith is our first step towards this mission.