From Widget Dashboard to AI Executive Assistant: The Full Story of Building alfred_
It Started With Widgets
In March 2025, the project that would become alfred_ didn't look anything like an executive assistant. It was called Boxento, a customizable widget dashboard built for students and productivity enthusiasts. The first commit landed on March 5, 2025, and it was simple: a grid of draggable widgets. Calendar. Weather. Notes. Todo. Quick links. World clocks. A Pomodoro timer. A currency converter. Even a geography quiz.
The idea was straightforward. People bounce between too many apps. What if you had one page that pulled everything together? The early days were pure frontend work: React, TypeScript, Tailwind CSS. We built a widget grid system where users could drag, resize, and arrange their workspace. We added authentication, sync so your layout followed you across devices, and dark mode because every developer tool needs dark mode.
Over the next few weeks, the widget library grew. RSS feeds. GitHub streak trackers. Year progress bars. Sticky notes with folders. The system was modular and clean. Each widget was its own world. Users could customize their dashboard to match how they worked.
But something was missing.
The Problem With Dashboards
A dashboard shows you information. It does not do anything with that information. You see your calendar events. Great. You still have to go to Google Calendar to respond to an invite. You see your emails, but you still need to open Gmail to reply. You see your tasks, but you manually created every one of them.
The dashboard was a mirror. It reflected your digital life back at you. But it did not reduce the work.
This realization hit us around the time we started thinking about who actually needed this product. Students liked the widgets, but students don't have the pain that makes them pay. The people who are truly drowning in administrative overhead are professionals: freelancers, consultants, attorneys, accountants, founders. People who bill their time. People doing two jobs: the work they are paid for, and all the administrative overhead that generates zero revenue.
A consultant making $200K a year spends roughly 30% of their time on admin. That is $60K worth of time spent on email triage, scheduling, follow-ups, and task management. A human executive assistant who handles all of that costs $60K or more per year. Only the wealthy get that leverage.
What if we could build that leverage as software?
The Pivot
The project was renamed from Boxento to alfred_ and moved under the ScribeHQ organization. The name came from two places: Alfred from Batman, the butler who handles everything behind the scenes, and Donna from Suits, the executive assistant who reads the room, handles forty-seven things while you are in a meeting, and only escalates the three that need your brain.
This was not a cosmetic rebrand. It was a fundamental shift in what the product was trying to be. We were no longer building a dashboard. We were building an assistant.
The widgets stayed, but their role changed. Instead of being the product, they became the interface through which the assistant communicates with you. The email widget is not just showing your emails. It shows you what your assistant handled, what it drafted for you, and what needs your decision. The calendar widget is not just a schedule. It shows you conflicts your assistant detected and meeting prep briefs it generated. The todo list is not manually maintained. It is populated from emails your assistant identified as containing action items.
The dashboard became the cockpit. The assistant became the copilot.
Building the Brain
In November 2025, we started building the backend agent system from scratch. The early architecture was rough. We built a routing system that would classify what the user was asking for, then hand off to a specialist: a todo agent, an email agent, a calendar agent, a notes agent, a web search agent.
Each specialist had its own set of tools: functions it could call to read emails, create calendar events, add tasks, and search the web. The first version had its share of growing pains. The router sometimes sent calendar questions to the todo agent. The agents would occasionally hallucinate tool parameters. The email agent could read emails but could not draft replies. We iterated fast.
By December 2025, we had email and calendar agents that could read Gmail and Outlook, list events, create drafts, and manage tasks. We adopted a proper reasoning framework and moved from simple routing to stateful agent handoffs, meaning one agent could pass context to another mid-conversation. This was the moment the system stopped feeling like a chatbot and started feeling like an assistant.
Donna Goes Autonomous
January 2026 was when everything changed.
Up to this point, the AI only worked when you talked to it. You would open the chat, ask it to check your email, and it would. But a real executive assistant does not wait for you to ask. A real EA is already checking your inbox at 6am, already drafting replies to routine messages, already flagging the three things that need your attention before you even sit down.
We built the Donna pipeline. It runs 24/7 in the background:
- The Email Poller checks your Gmail and Outlook inbox on a regular interval. It fetches new messages and feeds them into the pipeline.
- The Triage Engine classifies every email. Is this spam? Archive it silently. Is this a routine reply? Draft a response. Is this a meeting request? Check the calendar for conflicts. Is this something only the human can handle? Escalate it.
- The Worker executes the triage decisions. It generates drafts in your writing style, creates calendar invites, extracts tasks from email threads, and stages everything for your review.
- The Executor sends approved drafts, accepts calendar invites, and marks emails as handled.
The key insight was the autonomy tier system. Not everyone wants the same level of AI autonomy. We built four tiers inspired by corporate hierarchy:
- Associate. Junior assistant. Archives spam, drafts everything else for your approval.
- Manager. Trusted assistant. Handles routine replies and archives automatically. Drafts responses to important emails and flags external meeting requests.
- Director. Senior EA. Handles most emails and calendar items autonomously. Only escalates high-stakes decisions.
- Executive. Chief of staff. Full autonomy. Handles everything.
Users start at Manager and can adjust up or down. The system learns your preferences over time through a memory system that analyzes your sent emails to understand your writing style, your typical sign-off, and how formal you are with different contacts.
The Morning Briefing
With Donna running autonomously, we needed a way to show users what happened while they were away. We built the Daily Briefing, a morning summary that tells you:
- What Donna handled overnight.
- What is waiting for your decision.
- What meetings you have today, with prep notes for each one.
- What follow-ups are overdue.
- What tasks are due.
The briefing is the "escalation queue." It is the answer to: "Boss, I handled forty-seven things while you were sleeping. Here are the three that need your brain."
We also built a Meeting Prep system. Before every meeting on your calendar, Donna generates a brief: who is attending, what your last interactions with them were, what the meeting is about, and any relevant emails or tasks connected to it. You walk into every meeting prepared, without doing any of the preparation yourself.
Going Mobile: SMS and iMessage
In late January 2026, we asked a question. What if you did not need to open the app at all?
We integrated SMS and iMessage as a channel. You can text your assistant. "What's on my calendar today?" "Reply to Sarah's email and tell her Tuesday works." "Add a task to follow up with the client next week." "What did you handle while I was in my meeting?"
Building the SMS system was its own engineering challenge. SMS responses need to be 2-4 sentences. No markdown, no formatting, just plain text. The router needs to understand context from previous messages. "Move it to 4pm" means nothing without knowing you were just talking about a calendar event.
We built a separate agent system optimized for SMS constraints: specialist agents with dozens of tools, context-aware routing using a conversation window, and post-processing that strips any formatting and keeps responses concise.
The SMS integration lets you manage your autonomous assistant from your phone without opening any app. "What's pending?" shows items waiting for your approval. "Approve 1" executes the first pending action. "Be less autonomous" drops your tier. All via text message.
The Engineering Reality
Building this was not clean or linear. The bug trackers tell the real story.
We had a period where Donna's worker was silently dropping items because the autonomy filter was too restrictive. It only processed items it explicitly recognized instead of processing everything except what it should skip. Users would see emails come in and nothing would happen.
We had calendar events showing at the wrong time because of three compounding timezone bugs. UTC vs local time. A regex that matched "20" in "2026" instead of matching the hour. Wrong parsing order. Each bug was small on its own. Together they made every event show up at the wrong time.
We had SMS messages duplicating, with a single user message triggering seven identical responses, because of webhook retries hitting a function that had no idempotency checks.
We had the AI occasionally signing emails as "Donna" instead of as the user. We had email threads breaking because sent replies were missing the In-Reply-To header. We had the meeting prep system using cached data that was hours stale instead of calling the APIs directly.
Each of these bugs was a lesson in building systems that operate autonomously on behalf of real people. When an AI sends an email for you, it better get the timezone right. When it drafts a reply, it better sound like you. When it accepts a meeting, it better check for conflicts first.
We recently completed a major migration across multiple backend functions to move from cached calendar data to direct API calls. Every function now queries the source of truth in real time, handles token refresh with rotation persistence, normalizes UTC quirks across providers, and deduplicates events.
What We Built
Looking at what exists today, across roughly 800 commits over 11 months:
- A multi-provider productivity platform that connects to Gmail, Outlook, Google Calendar, and Outlook Calendar. Not just reading. Writing, sending, scheduling, and managing on your behalf.
- An autonomous AI assistant that runs 24/7, triaging your inbox, drafting replies in your voice, extracting tasks from emails, detecting calendar conflicts, generating meeting prep briefs, and learning your preferences over time.
- A tiered autonomy system that lets you control exactly how much the AI does on its own, from "draft everything for my approval" to "handle everything and just tell me what you did."
- An SMS and iMessage interface that lets you manage your entire professional life from a text message. Check your calendar. Reply to emails. Add tasks. Review what your assistant handled. Approve or reject pending actions.
- A widget-based dashboard that evolved from showing information to showing decisions. Every widget is a window into what your assistant is doing and what needs your attention.
The Philosophy
We do not compete on features. Gmail will always be better at email. Google Calendar will always be better at scheduling. But Google will never build the unified workflow because they are organizationally incapable of it. Separate teams, separate products, separate incentives.
We are not building a better hammer. We are building the workshop where all tools work together.
Superhuman helps you email faster. alfred_ emails for you. Motion schedules your calendar better. alfred_ manages your calendar, your email, and your tasks as one connected system. Reclaim protects your focus time. alfred_ eliminates the work that steals your focus in the first place.
The moat is not any single feature. The moat is the orchestration between them. Your meeting tomorrow relates to the email from last week. Your tasks came from conversations you had this morning. Your follow-ups are tracked because your assistant saw the commitment in the email thread. Everything is connected because your assistant sees your whole professional life, not just one slice of it.
What Comes Next
We are building toward a world where every professional has access to the same leverage that was previously reserved for executives with six-figure assistants. Where you wake up to a briefing instead of an inbox. Where your AI handles the forty-seven things that do not need your brain so you can focus on the three that do.
The widget dashboard was version one. The autonomous assistant is version two. The vision is version three: one avatar, one interface, all your information. A digital presence that knows you (your work, your relationships, your goals) and grows with you over time.
We started with weather widgets and Pomodoro timers. We are building the future of work.
alfred_ is the executive assistant everyone deserves but could not afford. Until now.