Executive Assistant Agent
An agent that runs an inbox, a calendar, and a daily briefing.
Overview
An AI agent that handles the admin around a founder's day. It sorts the inbox, drafts replies that sound like you, books meetings across time zones, and puts together a morning briefing from your mail, calendar, and docs. Instead of jumping between four tools, you talk to one assistant in plain English.
The problem
Founders lose an hour or two a day to sorting email, playing calendar Tetris, and chasing down where things stand. The context you need to answer one email is spread across Gmail, Calendar, and Drive, so even a simple request means opening five tabs.
The assistants you can buy off the shelf either can't take real action, or they ask you to hand over full access with no guardrails. What was missing was something that could actually operate the tools, but only inside limits you can see and check.
The solution
I built an agent that plans before it acts. It reads a request, decides which tools it needs, runs them, and checks the result before it reports back. Each thing it can do (read the inbox, draft a reply, find a slot, summarise a thread) is a separate tool with its own rules and permissions.
It also remembers. A retrieval layer keeps track of past threads and your preferences, so replies match your tone and the recurring decisions get handled the same way each time. Everything it does is logged, and anything it can't undo, like sending or deleting, waits for you to say yes.
Architecture
An orchestration layer runs the plan-act-check loop over a set of typed tools, backed by retrieval memory and a human approval step for anything sensitive.
- Each tool has its own access scope and rate limits
- The agent checks its own work before it writes anything externally
- Retrieval memory for tone and recurring decisions
- Everything is logged, and anything irreversible needs a yes from you
Screens & diagrams
Key features
Sorts and drafts your inbox
Sorts what comes in, flags what needs you, and drafts replies in your voice with the right context attached.
Books across time zones
Reads availability, offers slots, and books the meeting, including the reschedules, without the back and forth.
Morning briefing
One short digest of the day's meetings, the threads that matter, and open items, pulled from mail, calendar, and docs.
Guardrails you can see
Anything it can't take back waits for your confirmation, so you can trust it with real actions without letting it run loose.
Implementation highlights
- Made each tool a self-contained piece, so I can add a new capability without touching the planning loop.
- Put explicit timeouts and backoff with jitter on every external call, so rate limits and outages don't take it down.
- Wrote a memory summariser that keeps context inside the token budget while holding on to past decisions.
- Built a test harness that replays real requests, so I catch regressions before they ship.
Challenges solved
Trusting it without letting it loose
The approval step only intercepts the actions you can't undo. Everything else runs freely, so the agent stays fast but never does something you can't take back.
Replies that sound like you
Generic drafts feel like a robot wrote them. Pulling from your own sent mail, plus a short style profile, got the drafts close enough that they go out with barely an edit.
Staying under the rate limits
Google's and the model provider's limits meant I had to batch, cache, and back off carefully so the agent stays responsive without tripping quotas.
Technologies
Future improvements
- A voice mode for hands-free briefings and dictated replies
- Surfacing follow-ups before you have to ask
- A team version with shared context and per-person permissions