Personal AI agents have owned the tech zeitgeist for the past 30 days. LLMs are incredibly powerful, but mainstream consumers are using AI chatbots as glorified search engines. Instead, the promise of these agents is simple: personal AI agents should transform the coordination and execution of life operations.

Meta's Muse has emerged as the early leader in personal AI agents. It's a thoughtfully designed product with the right mix of practicality, play, and platform capabilities that could allow Meta to usurp ChatGPT and Claude as the leading consumer application for AI.

Naturally, I've spent an unhealthy amount of time testing Muse to its limits: asking it to negotiate a car lease and swap insurance, manage my fantasy football team, make restaurant reservations, and manage the distribution of my Raptors season tickets. It's been met with varying degrees of success, and while the short-term value is mixed, the long-term vision is profound. The most illuminating part of my two-week experiment with Muse was actually the initial sign-up, where Muse articulates its value in three central tenets:

Meta, in outlining the three central tenets of Muse, is actually outlining the three central tenets of all AI agents, consumer and enterprise. For hundreds of millions of people, Muse will be the product that ingrains agentic behaviour. In doing so, Meta is setting the capability and human-computer interaction bar for all AI agents.

Here is how we're approaching these tenets at Gigi:

Can take action for you

This is the foundational value proposition of Gigi, and it should be the foundational value proposition of all agents. Customers need to trust that their AI agents can take action for them. They need to know the breadth of capabilities (for us, this means all the capabilities of the Amazon DSP) meets the depth of capabilities (that they are executed reliably, accurately, and fast). Our customers feel the magic of working with Gigi once they know they can transform their way of working from laboriously pressing buttons to assigning work to Gigi and orchestrating her to complete it as optimally as possible. Once they see the glimmer of magic, they naturally reward Gigi with more work. Across our customer base in September, Gigi completed just under 800,000 actions (+313% from August), 41% of which were automated. That level of scale and growth is earned by the trust we've built with our customers, one approved action at a time.

Works around the clock

If agents only take action on your behalf when you prompt them, they are merely co-pilots. A team member shouldn't only work on the problem they own when you prompt them, and neither should your agents. At Gigi, the primary mechanism our customers use to interact with our product is not ad hoc prompts but ongoing tasks. Gigi then works on these tasks on a cadence the user defines, or seeks out opportunities every day to achieve the task at hand.

For both Muse and Gigi, this aspect of the agentic experience could use some work. I'm increasingly finding Muse's notifications to be noisy and not at the fidelity I require. At Gigi, tasks run every single day, and while that's been working thus far, we know where we can improve. The daily cadence can, like Muse, be noisy for our customers. Gigi can at times provide too much information and too many actions to review, and we're actively questioning whether certain aspects of a task need to run daily (at minimum, this would improve our token costs). The daily cadence can also mistake movement for impact: too many little actions don't leave the cooling period needed to measure their effect and draw a clearer line to causality. Despite these known areas of improvement, knowing that Gigi is "always on" and doing the work of our customers is a clear distinction of value that's allowed us to succeed.

Stay in control

Control and trust go hand in hand. If a customer doesn't have a crystal clear grasp of the actions their agents are taking, they will never defer to agents for more frequent or more complex actions, and the opportunity to provide value will be lost. We've designed this very intentionally at Gigi. By default, Gigi does not take any action without human approval. Only after trust is gained does Gigi begin to proactively suggest automating certain actions so that humans do not become the bottleneck for value. And even when actions are automated, we've invested extensively in an audit trail in which our customers have full visibility on who took the action (Gigi or human), when the action was taken, who delegated the action to Gigi (if automated), and a clean process to reverse any action taken. We've invested a non-trivial amount of engineering resources in enterprise guardrails, permissions, and administrative controls because we're cognizant that we only win when there is the utmost trust in Gigi.

Which brings me back to that sign-up screen.

Meta is about to teach hundreds of millions of people what an agent should feel like. Those people are our customers' employees, our customers' clients, and our customers. They won't grade Gigi, or any enterprise agent, against the legacy SaaS they used last year. They'll grade her against Muse. Meta is onboarding our future users for us, and setting a bar we know we can meet, because it's the same bar we've held ourselves to for two years.