Our Master Plan

Summary

We are on a mission to automate the tasks that currently demand our time but not our intellect. We envisage a future in which we each have a team of “remote” workers to which we can delegate those tasks, and we’re building a new platform to bring that future to a reality.

Platforms are the foundations onto which other applications are built. The computer or the smart phone provided a user interface, memory systems, networking hardware, compute and a kernel to application developers, who were then free to focus on building great applications.

Today, Agents are the new applications, but developers seem to be building the same components over and over again: reasoning capabilities (models), complex memory systems, preference management frameworks, authorization & authentication suites, temporal awareness and completely new user interfaces.

We’ve integrated all of these components into our platform, PersonaOS, and given it the interface of a “remote” human employee - with each Persona having it’s own identity, email address and phone number. We are making all of this available so developers can focus on building great Agents.

Like the developers of the personal computer and the smart phone, we recognize that we need to provide a core set of applications to drive initial user adoption, like the initial “Home Screen” apps on the iPhone. For that reason, we’re also building a suite of applications logically belonging to an executive assistant, including scheduling, expenses and reminders.

Join us as we build the future onto which the next century is built.

Mission

Our goal is to automate all tasks that require a human’s time but not their judgement.

It’s said that no decision gets to a President’s desk unless the choice is “51-49” - 51% in favor of one option, 49% in favor of the other. All day, the President makes judgement calls on which option is the 51% and which is the 49%. From the Chief of Staff to the Secretaries, a team of people works to filter out all of the “60-40” and the “70-30” decisions, but unlike the President, most of us don’t have a team and instead spend our days doing tasks that are “99-1”.

We do tasks like filling out expense reports (deciding which expense category “Coffee” goes into), scheduling meetings with clients (deciding if 8 am is too early) or booking travel (checking if the aisle is available). We don’t have a team of people that understands our preferences and can make those judgement calls on our behalf - and so we make them ourselves.

Our mission is to enable everyone to make only the “51-49” decisions in their lives. Today, we are automating the “99-1” tasks - like expenses, scheduling and travel - and tomorrow we’ll unlock the “98-2”. Then we’ll tackle “97-3” tasks, and so on.

Vision

“Remote” workers for you and your team.

Most of us are stuck making “99-1” decisions because we don’t have a team of people to make them for us and because existing software doesn’t replicate our own judgement well enough to make the decisions for us.

Incredibly, foundation models can now approximate human judgement, and we can provide enough detail about our own lives, habits, and tendencies for these models to simulate our personal preferences. For the first time, the opportunity exists to create software the can make judgement calls for us as individuals.

We believe the best way for software to capture these preferences is through natural, human-to-human interaction. The future we’re building includes billions of “remote” coworkers and employees which interact with you as you’d expect any other remote human colleague to: over phone, email, text, Slack and many more applications.

Through instructions or feedback on calls and messages, your “remote” coworker derives your preferences and thereby better aligns its judgement calls with your own. This feedback loop is innately human: the first time you ask your “remote” coworker to book a flight, they might incorrectly book you in the window seat. But with your feedback sent over email or spoken over the phone, the coworker learns to get it right the second time (and the third) by remembering your preferences.

The future is filled with software that learns from us as we expect other humans to learn from us. We understand that this seems far fetched and as if we’re selling “AGI”, but we’re not. Read “Platforms & Agents as Applications” below for our plan.

Platforms & Agents as Applications

Background

The term “platform” is entirely overused in the startup community, but bear with us.

Taking a step back, a software platform is a foundation on which other applications can be developed, and there have arguably only been two software “platforms” to date: the computer and the smart phone. Both of these devices provide application developers with:

1. User interfaces for users to interact with applications,
2. Memory (both RAM and Disk) to store information,
3. Networking for applications to access the internet,
4. Compute to do calculations based on user inputs, and a
5. Kernel to manage the concurrent execution of multiple applications.

A New Platform

In the context of automating the “99-1” decision for humans, the existing “operating system” platforms are insufficient. Human judgement and contextual understanding drove operating system design, and adding foundational models to everyone’s laptop is similar to adding a combustion engine to a horse drawn carriage while keeping the horse. It may lighten the load for the horse in the short term, but eventually, the combustion engine-powered carriage will simply be dragging the horse behind it.

ChatGPT Image Jun 8, 2025, 04:14:55 PM

To further illustrate the point, consider all of the components required to create a platform that usefully simulates human judgement:

1. Reasoning capabilities (from foundational models),
2. Memory of past interactions (not to be confused with RAM or Disk),
3. Preference prioritization algorithms to know when to follow what preference,
4. Authorization (and Authentication) to use the apps we humans already interact with,
5. Temporal understanding so decisions occur when they’re relevant, and a
6. User interface that can be used for many different types of tasks.

And this doesn’t even include the execution of the task after the decision has been made. To combine everything, a new platform is needed to allow developers to focus on building high quality Agents.

That said, we don’t want to separate users from their existing platforms - computers and smart phones are incredibly good at serving information to humans (the horses from our prior analogy). Instead, we have created a “virtual human” that has a consistent identity and memory across all of your existing communication tools. This “interface” enables humans to delegate tasks naturally, as they would to another human.

We call this interface a “Persona” and the system that runs it the “PersonaOS”.

Within the PersonaOS, we’ve built a proprietary memory architecture to structure memories specifically for Agents. We have developed a means of not only authorizing Personas to do tasks on behalf of users but also intelligently storing and retrieving preferences for how do those tasks. Personas have a complete understanding of “time”, with the ability to wait for, re-awake on and remember at specific times and dates. And better still, the human user interacts with a Persona using all of the tools the user already uses.

Agents as Applications

In this new platform, we define an “AI Agent” as an application.

An AI Agent is a set of tools, prompts and models that combine to complete a logically similar set of tasks.

Within the PersonaOS, an Agent is simply an application that is used when the Persona identifies that it needs to complete a task that an Agent is suited for. For example, if a user asks their Persona to schedule time for lunch with a colleague, the Persona will use the “Scheduling” Agent which is designed specifically for scheduling (and rescheduling) meetings on behalf of the user.

When users create a Persona, they can choose which Agents (applications) it has access to, and thereby choose the extent of the capabilities of the Persona. This is conceptually similar to your own capabilities being limited by the applications you have on your laptop.

For technical readers, all AI Agents in our platform are containerized using Docker images. For Agent developers, PersonaOS resources such as memories, preferences and user interfaces are accessed through Agent-ready SDKs that we provide.

At the start (discussed more below), we will provide a small set of Agent applications by default to encourage user adoption, but we intend to create an AI App store, in which developers can upload their Agent for approval and then naturally extend the capabilities of Personas.

Minimum Applications

When launched, Windows 1.0 and iPhoneOS offered immense value of the platform to application developers, as described earlier. However, neither the computer nor the smart phone was able to launch without a small set of applications that were immediately useful to the user.

A minimum set of applications drove user adoption before developers were willing to begin building, and neither of these platforms had an “App Store” equivalent at launch.

We don’t have to provide a fully built Agent App Store from day one - in fact we believe that it would be foolish to spend time on it before we scale. We need to focus on building out the PersonaOS itself and the minimum set of Agents that make the platform immediately useful for any user who interacts with it.

Starting Small

The “Executive Assistant” Persona contains the minimum set of applications to launch the PersonaOS.

Unlike the computer or the smart phone, we don’t have to sell hardware to drive adoption of PersonaOS. Instead, our goal is to create the minimum set of applications required for user adoption - and that starts with a phenomenal executive assistant.

Executive Assistants

Shockingly few people understand what an executive assistant actually does and what the difference between executive assistants and administrative assistants is. That’s why we interviewed executive assistants to find out what they do.

Administrative assistants largely help with tactical, day-to-day tasks such as scheduling meetings, doing expenses, booking travel or maintaining filed documents. From our customers, this is what we believe most people imagine an EA does, but the role of an EA is much larger than that.

A great executive assistant is (1) a guardian of time and (2) a project manager. While most executive assistants we’ve spoken to began their careers as administrative assistants and continue to do those tasks, the real value they provide as an EA is as a buffer and manager for their executive. To illustrate this, consider the following (real) examples:

"If you ask five people in the company, everyone thinks their stuff is the most important for the CEO to listen to. When in reality, that’s not the case.”

"We have two major projects going on right now... I can sit in and if I hear a red flag that she [Executive] needs to be aware of, then I'm like, okay. She needs to be aware. She needs to be in the next meeting. We can't move forward without it.”

"I had enough knowledge of my relationship with my executive and with the priority and everything that I was like, no. This one cannot move forward without him. We need to wait and find a time.”

In all of these examples, the EAs were exercising judgement based on the executive’s preferences and priorities. In other words, the crucial difference between a tactical administrative assistant and an effective executive assistant is an understanding of how to operate as an extension of the executive.

Initial Applications

To build out the “executive assistant” applications within the PersonaOS, we are starting with the tasks of an administrative assistant: scheduling, expenses and reminders & recurring tasks. Over time, we will continue to expand project management capabilities within the PersonaOS.

Scheduling

Scheduling is both an incredibly easy and an incredibly difficult task. For setting up an initial time, Calendly (valued at $3bn) or Cal.com is effective and easy to use - our Scheduling Agent can do that, but that’s not what we’re focused on because that’s the easy part.

We are focused on the days when you need to clear your afternoon, your friend is trying to find a dinner time with you or the board meeting with 10 people needs to be moved. Our Scheduling Agent is designed to make maximum use of your expressed preferences and understand how to operate when the ideal outcome isn’t clear.

Calendly doesn’t understand the difference between “find some time” with a colleague versus with a friend - our Persona Scheduling Agent does. For example, you may say “don’t schedule things on weekends” and also say “John is a friend, we can meet whenever.” For your friend John and all other friends, you would expect the preference to be “meeting on weekends are allowed.”

For coordinating with multiple people, the Persona is perfectly suited to act as if it were just another human assistant - sending and replying to emails with multiple people. The PersonaOS handles waiting on future emails, mapping new emails (even in different threads) to the existing scheduling task and even the timing of follow-ups.

Currently, the Scheduling Agent fits between the administrative assistant and the executive assistant. It’s capable of advanced scheduling operations and learns over time from the user how they prefer things are done.

Expenses

Everyone hates doing expenses, but it’s a necessary evil that isn’t going away anytime soon. For that reason, it’s one of the first things given to EAs to get done.

While many companies are switching to Ramp (recently valued at $13bn) or Brex, many companies will remain for awhile on SAP Concur, Expensify, Chrome River, Coupa Expense, Oracle EM and Workday Expenses. We have yet to talk to someone using one of these platforms that does not become immediately excited when we say we are working on an Agent that looks through your inbox and calendar to find and do your expenses. In that, we have the opportunity to create a Ramp-like experience for expenses for everyone else.

Reminders & Recurring Tasks

Out of the box, PersonaOS can manage reminders and complete tasks on a recurring basis for users.

For reminders, users can ask for either one off or recurring reminders that are sent however the user specifies them. A user can ask to be emailed tomorrow at noon or called once a day with a breakdown of their inbox and action items.

For recurring tasks, users similarly ask in natural language, such as “every morning, send me a briefing on today’s calendar.” Each time the task needs to be executed, the PersonaOS handles the assignment of the task to the correct Agent App to get it done.

Future Applications (Examples)

While we’re starting small, we’d like to emphasize the platform opportunity with a few examples. Each of the following domains is a natural Agent App within the PersonaOS, but they are all examples that we will not build ourselves.

Coding

Coding Agents are increasingly interacting with users outside of the IDE as they attempt to move towards a “SWE-as-a-Service” model. The most recent example being Cursor’s Slack integration.

Inside PersonaOS, a developer could build a coding Agent to interact asynchronously on their codebase easily. PersonaOS provides the authentication into Github, the understanding of user preferences for code styles and the interface for interacting with users. All the Agent developer has to do is build the coding Agent itself.

For an example of how easy this is within PersonaOS, we can put Claude Code (with a FastAPI wrapper) into a Docker image and then ask a Persona to do work on our codebase. In this example, we’ve created a coding application with in PersonaOS using an existing, powerful coding Agent and little more than a Dockerfile and a single FastAPI file.

Deep Research w/ Browser Use

We are often asked to help with some version of the following: “I need to evaluate [N] options for [Y] event with a specific filter and then make a decision on which is best.” A few specific examples:

• Finding a nice table at a restaurant, filtering for availability, noise and quality
• Evaluating wedding venues, filtering for availability and cost
• Rebooking flights or hotels, filtering for timing and cost

These tasks typically require the use of browser to fill in forms and a phone to call the vendors, along with the ability to wait and follow up with each when appropriate. The PersonaOS already handles the project management, follow-up triggers and phone call infrastructure required for this process, and a developer could simply add an Agent that specifically designed for browser use to be able to accomplish this set of tasks.

Financial Diligence

Financial diligence usually contains a consistent set of artifacts that need to be created at each stage of the investment. In private equity, this may include a model (like a DCF or LBO) and a deck detailing the assumptions and market research.

As an example, an Agent developer building a “DCF Agent” would need to build a user interface, the integrations to the team’s shared drive for data room access, the Agent itself, the coordination protocols for processing feedback from multiple investors, the preferences architecture to remember how each specific investor prefers their excels formatted and the Agent to actually build the DCF itself.

Instead, within the PersonaOS platform, the Agent developer can simply build the Agent to build the DCF itself. The PersonaOS already provides the preference management, the user interface, the coordination with multiple individuals and the authentication required to access company documents.

Team

Julius Stener

Julius received his BS and MS at Stanford, where he studied Computer Science focused on machine learning. He also previously worked at Activant, a growth stage venture firm in NYC, and for the Department of Defense at US Indo-Pacific Command.

Joseph Matan

Joseph received his BS at Stanford, where he studied Mathematical & Computational Sciences. He has spent the last 4 years implementing AI inside enterprises - designing Agents to replicate human-level knowledge work in finance.

The Team

Julius and Joseph rowed together and were roommates at Stanford. After Stanford, they both lived in New York and had many late night white-board sessions working on the architecture of the original PersonaOS, and when Joseph moved to LA in 2024, they continued to talk often about the solution.

Frequently Asked Questions

Hasn’t someone tried this before?

People have - many times! The three most notable attempts are Fin Assistant, X.ai Assistant and Clara Labs.

Fin Assistant was cofounded by Sam Lessin and Andrew Kortina in 2015. They both wrote post-mortems (here and here), which have been quite helpful to understand the reasons it didn’t work. Notably, they attempted this prior to useful foundational models, and built software that intelligently routed the tasks to humans who focused on that subtask all day.

Based on the post-mortems, the primary reasons that Fin didn’t work were (1) the completion accuracy wasn’t high enough and (2) the financial cost of incorrect task completion was too high. The latter occurs when the EA is actually spending money (booking flights or buying flowers) on behalf of the user, and both are driven by how accurate the assistant is.

Founded in 2014, Clara Labs has recently relaunched with their scheduling assistant: here; however, they focus exclusively on scheduling today. Originally, they used significant human labor as well.

Is someone trying this now?

Of course! There are different types of executive assistant startups.

Howie does not offer personalized assistants - only a single assistant persona named howie (i.e. "howie@howie.ai"). This design choice limits their ability to reach a level of personalization that this space needs. Personalization (i.e. 1+ EA "identities" per user) is required because of the permissioning challenge that exists here. For a virtual EA to do its job well, it need to understand when someone has authority to do something and when they don't - if two users email the same "howie" in the same thread, deciding the permissions for the Agent that actually handles that is a nightmare. And that's for scheduling tasks - it will be much worse for higher complexity / risk tasks.

Workmate is probably the best competitor on the market. We’ve used their EA, and it suffers from the same issues that our current EA does (namely hallucinating capabilities). They have a better UI/UX for the onboarding flow, and we are currently working on ours. The two big differences in our philosophies are this: (1) they have a chat app, we do not and (2) they have a detailed "settings" page.

On point one, we don't think an EA should never break the barrier into "software", and instead it should be fully native to our existing human communications systems. That means you can call and text it like a human, but you can't login to an app on your phone and chat with it.

On the second point, we believe that the human should never have to explicitly define "preferences" in a settings page or a custom user-defined system prompt. To avoid that, we extract "preferences" from user messages and store them in our knowledge graph - so instead of having a settings page that says "don't send on weekends", our users just text their EA and tell it "don't schedule on weekends." It achieves the same thing, but our method is far more extensible. Because "preferences" are just strings, remembered from previous interactions and populated at Agent runtime, they can be complex and don't need to fit into a rule / data schema. It allows you to do things like write "don't schedule meetings on weekends" in one email and "it's ok to schedule meetings with Jack on weekends, he's a friend" in another.

Why won’t OpenAI, Anthropic or Meta win this?

Building a model is fundamentally different than understanding, capturing, storing and recalling user preferences for Agents. This is explored above in “A New Platform”, but reasoning is only one component of accurately completing tasks on behalf of the user. Of the remaining parts, the most difficult (and differentiated) component is preference capture and prioritization for Agents doing work on behalf of the user. OpenAI, Anthropic and Meta are not focused on this today.

To illustrate the point, we’ll highlight a few problems you need to solve in order to successfully build what we’ve built that are not solved by traditional memory techniques or simply integrating to knowledge stores:

1. (De-)Conflicting Preferences

A user can simultaneously say “I don’t want to take meeting weekends” and “I want you to book a date night with my wife this Saturday.” In this context, the assistant should be more than capable of scheduling a date night on the weekend, despite explicitly violating a stated preference.

Our system loads preferences based on the context of the situation to ensure that conflicting preferences are handled as expected.

1. Similar Names / Titles

Assume a user knows multiple people named “John”. The user has previously told their assistant that “John loves fine wines” when referring to one John and said “John doesn’t drink” when referring to another John. If the user asks to plan an evening event, traditional vector-based retrieval would not work - as both would be retrieved. Graphs are required.

1. Post-Authentication / Post-Authorization Permissions

Let’s say two users, Bob and Sue, work with the same assistant, Jane. Bob has given Jane access to his email account so that it can schedule for him. Despite Jane having authorization and the authenticated credentials to access Bob’s email, Sue should not have access to Bob’s email through Jane. Jane must intelligently determine user permissions during execution. This is a permission post-authentication and post-authorization.

1. Topic Recall

You expect to be able to talk to your assistant about a topic today and then resume the conversation months from now. Further, you expect that your assistant remembers the prior conversation perfectly and can pick up where you left off.

1. Asynchronous Information

You must be able to interrupt your assistant mid-way through task execution with new information.

The above problems (and more) exist for anyone building a “Persona,” and these are not currently being solved by the model providers. That isn’t to say they won’t focus on them in the future, but for now, our understanding is that they are focused on bringing integrations into their chatbots UIs.

Why won’t Apple or Microsoft win this?

Apple and Microsoft’s business models are dependent on the sale of hardware devices and software licenses, respectively. In Apple’s case, 75% of revenue is hardware, and software licenses are roughly 69% of revenue for Microsoft.

As a result, both have and will continue to bring AI into their ecosystems in such a way that it doesn’t cannibalize their existing revenue streams, meaning they will continue to try to tie an assistant to the device or the license itself.

However, “virtual” AI employees are worse products when tied to a single device or license - it prevents them from existing everywhere and across all systems. It also presents a challenge for Apple and Microsoft when selling AI employees independently of their existing revenue streams (i.e. buy an employee on it’s own).

Why won’t Google win this?

Google has our memories and preferences (in email and to a lesser extent Google Meet). Their business model is not dependent on hardware or license sales (except for small parts of Android). They have the frontier models in terms of cost-accuracy trade off. So why won’t they win?

They might, but they would have to enter and succeed in a market with significantly lower margins than the one they currently operate in - as a public company. They would need to develop their own PersonaOS equivalent and then thoroughly test it before public launch, and they would need to do this without losing the employees working on it (as they go do it themselves). But they might get it done.

Why are you the team to do this?

Julius spent the last year building 3 versions of the PersonaOS before finally learning enough about the pitfalls involved to make one that worked. Joseph spent the last few years deploying Agents at scale inside companies as both a consultant and as an in-house developer. Both experiences mean that there are actually very few people who understand the problem and the solution better than they do.

Also, they both rowed for a decade each - sitting in a chair coding or selling the product is a lot easier than getting up at 5 am to freeze and hurt everyday.

What are you working on today?

Engineering - we’re focused on (1) reducing context-driven errors and (2) implementing Zoom and Outlook integrations. That means we want to make sure Agents are only accessing the correct memories and preferences for the situation and at the correct time. Currently, the scheduling Agent is roughly 85% effective at rescheduling, dropping the ball or otherwise appearing non-human.

Sales - we are working on developing the value equation with customers. This will naturally evolve over time as we add more Agent applications to the Persona.

Product - after feedback from users, we are redesigning the onboarding flow to better capture user preferences up front. The redesign will be similar to how you would expect a human EA to ask you questions about your preferences on their first day.

How much are you raising?

Let’s talk about it. Email us.

What will you use the money for?

We will launch with our first enterprise customer in Q3 and our next customers in Q4. To achieve this, we will be rebuilding our onboarding workflow, expanding our integration set (including with Outlook, Zoom and SAP Concur), adding enterprise features (such as custom domains and administrative accounts) and improving the accuracy of our Scheduling Agent.

We have a candidates already identified for both engineering and design positions to enable us to hit the ground running after funding.

What about Security and Privacy?

We are currently in the process of becoming SOC II compliant with Vanta. We expect to receive our SOC-II compliance document in Q4 2025.

How big is the market?

$200bn - Existing US Human Assistant Market

The existing market for human administrative and executive assistants in the US today is larger than $200 billion per year. That is 3 million assistants with an average all-in cost of $70 thousand per year.

$50bn - Virtual AI Assistant Market in 2025

If every knowledge worker in North America and Europe (~150mn) paid for an AI assistant at ~$330 per year, the total addressable market for AI assistants would be roughly $50bn. However, note that this is a gross underestimate because the value provided (and the value we charge today) is significantly greater than $330 per year.

>$1tn - Global AI Knowledge Worker Market in 2035

In 10 years, AI knowledge workers will be present across all enterprises, and each consumer will have an assistant just like they have smart phone. That future includes billions of AI workers.

What is your business model?

We are starting with a subscription model because that is what customers are used to; however, the natural business model for this product is usage based - similar to human employees. The result will likely look closer to how we pay for human labor today, with graduated levels depending on the perceived number of hours worked. Some people only need a part-time assistant while others need a full-time employee. Others want unlimited hours of work available to them. We will benchmark off of “tasks completed” to determine how much each level is prepared to do before the user is notified that they are paying for overtime (i.e. additional usage beyond their subscription).

Who is your target customer and how are you going to acquire them?

Our initial target customers are the 23-30 year old investment bankers, consultants, lawyers and buy-side professionals. From our customer discovery, this group of people has direct exposure to the benefits of EAs, but they are not senior enough yet for the firm to pay for them to have EAs.

For our sales motion, we plan to sell to the COO's office and focus on highlighting the reduced cost of doing administrative tasks manually (and the increased productivity). We are beginning this shortly.