Sync

Aims to ease the daily booking ritual, and return the space nobody is using.

SYNC AGENT
Sync welcome screen on an angled phone
Sync home screen with the day already booked
Desk H2-12 · near DesignBooked 3:30 PM, before you asked
Parking P12 · Level B1You drive on Tuesdays

Sync · concept + interactive prototype

What if your office booked itself around your day?

A booking assistant for hybrid offices: plan the desk, room and parking around a changing workday. I designed when it acts, when it asks, and how people stay in control. Agent behavior and booking data are simulated in the prototype.

Responsibilities
Desk research · UX/UI design · Prototyping
Timeline
1-day sprint
Industry
Workplace technology
Three Sync screens: Atlas is back in the pool, Home with the day booked, and Booking your Tuesday

The context

Modern offices increasingly run on hot-desking, where employees no longer own a fixed seat. Hybrid workers come in only on some days, mainly for meetings and collaboration, and have to book whatever they need each time.

The challenge

Hot-desking freed the floor plan but pushed the work onto people. Every day meant re-booking a desk, chasing rooms that vanished when a meeting moved, and living with ghost bookings that held space nobody used.

  1. A tired employee juggling a calendar, a desk map and a parking list on floating screens

    Manual, repetitive booking every single morning

  2. A meeting slot turning red and moving on a floating schedule inside a glass meeting room

    Rooms disappearing the moment a meeting shifts

  3. An empty office desk still marked as booked, with faint outlines of people who never came

    Ghost bookings holding space others are waiting on

Brief

Design an AI assistant that handles all of this ahead of time, while keeping people firmly in control: actions explained, correction available, and autonomy bounded by the limits they set.

My approach

Before shaping the interactions, I reviewed Microsoft’s Agent UX principles, its HAX guidance on correction, and Anthropic’s Building effective agents. I used them to define three decisions: make background work visible, keep correction close to the action, and put explicit checkpoints around autonomy.

A booking form asks people to choose every resource. An agent asks them to delegate a goal, then needs to show what it may do, what changed, and when it needs help. I designed those boundaries before polishing the screens.

The guidance behind the design decisions

01

Make background work visible

Microsoft connects trust with transparency and control. I translated that into live task steps, “Why this desk” and an Activity record.

UX design for agents ↗
02

Make correction part of the flow

HAX recommends making AI mistakes easy to correct. I kept edit, pause and manual takeover close to each action, with alternatives when Undo cannot restore a space.

HAX · Support efficient correction ↗
03

Separate planning from permission

Anthropic describes agents checking tool results and returning to people at checkpoints. I applied that to a plan → check → act loop, with confirm-first and exceptions in Full auto.

AnthropicBuilding effective agents ↗

These are my applications of the guidance to Sync. The sources informed the concept; they do not validate its usability or reliability.

My approach, as a FigJam board. Understanding the context and user: what is hot-desking, how is it managed, who are the users, what does their day look like. Identifying friction points: what is the friction, when does it occur, how does it affect people. Ideating on agent interaction: what can the agent do, how much control does it get, when does it act. Connecting business needs and user goals: efficient use of resources is efficient use of time. Designing the solution: a scalable design system, human in the loop, a context-aware agent, privacy and security controls, and an easy manual override.
From understanding the context and the friction, to how much the agent should do, to the business need behind it.

Design strategy

Business objective

Use meeting rooms, desks and parking efficiently, take the friction out of booking, and support a smooth hybrid work culture.

Solution

An AI-powered app that works quietly in the background and takes the load off employees.

Goals

  1. Remove repetitive manual booking
  2. Free up “ghost-booked” spaces for others
  3. Make better use of shared desks, rooms and parking
  • Reads context

    Uses your calendar, location and past booking preferences.

  • Re-routes automatically

    Proposes or rebooks when plans change, within your permissions.

  • Books ahead

    Plans your office days and books what you'll need.

Who is Sync for?

Concept personas based on desk research, not interview participants. Their quotes express the needs explored in this concept.

Portrait of Neha

Neha

Product Manager

Demographics

Age
28
Work style
Highly collaborative, constantly switching between focus work and meetings.
About
Curious and adaptable product manager who enjoys solving complex problems and working closely with different teams.

“My day changes by the hour, so my bookings should keep up without me babysitting them.”

Pain points

  • Constant manual updates to existing bookings
  • Filtering for the right room takes effort
  • Different resources for each meeting

Needs and goals

  • Meeting rooms with video-conferencing
  • Sitting near different teams depending on the day
  • Context-specific spaces, like a room with a large whiteboard
Portrait of Kabir

Kabir

Senior Data Analyst

Demographics

Age
29
Work style
Deep-focus worker, in the office 1 to 2 days a week, prefers a quiet environment.
About
A methodical analyst who does his best thinking in long, uninterrupted blocks and plans his office days around the setup he needs.

“I drive in for deep-focus days, and if my setup and parking aren't guaranteed, the trip is wasted.”

Pain points

  • Struggles to find parking
  • Misses his equipment setup when he forgets to book 48 hours ahead
  • Doesn't always get a quiet space

Needs and goals

  • A specific setup, like dual monitors
  • Quiet zones for concentration
  • A guaranteed parking spot for his long commute

User journey map

A desk-research journey hypothesis: from planning to the moment plans change. Each stage points to an opportunity for Sync, ready to test with users.

User journey map: four stages of a booking day
Plan the dayBook workspaceBook supporting needsChanges happen
User actionsChecks meetings, office location and timingsOpens the booking platform and searches for a deskBooks parking and meeting rooms separatelyChanges or cancels bookings by hand
EmotionsNeutralFrustratedOverwhelmedAnnoyed
Pain pointsInformation is spread across calendar and workplace toolsManual search and selectionSeveral tools and separate bookingsEvery change means more manual work
OpportunitiesUnderstand the user's whole workdayRecommend the right workspaceBundle desk, room and parkingAdapt bookings to changes automatically

Ideas before interfaces

Ideation notes: how an AI agent could use the calendar and past habits to remove manual booking, and two concept sketches, a summary card with Change and Keep, and a calendar showing an agent re-route.
Early notes and two first concepts: a summary card, and an edited calendar.
Two home screen sketches: one where a delayed meeting gets a new room with Undo and Edit, and one with a morning summary of the desk, room and preferences, plus a + button for manual booking.
Home screen sketches: the agent quietly re-routing, and a daily summary with clear controls.

Meet Sync

A concept for an assistant that plans your spaces, explains its choices, and acts within the permissions you set.

  • Takes away the mental load

    When plans change, Sync proposes or handles rebooking within your chosen limits.

  • Cuts ghost bookings

    Forgotten bookings are flagged as soon as a meeting moves or is cancelled, and released once you agree.

  • Keeps you in control

    Actions stay visible, with controls to undo, edit or take over.

Agent interactions

Design decision · Following Microsoft’s transparency and control principles ↗, I gave autonomy, status and memory their own visible controls.

Three screens set up the relationship with the assistant: how much it may do on its own, what it’s doing right now, and what it remembers about you.

  1. Autonomy level screen: How much should I decide alone? Full auto, Confirm first (selected) and Manual

    1 of 3: You set how much it does

    From fully automatic to confirm-first, separately for desks, rooms and parking.

  2. Agent screen: Lining up your week, reading calendar, finding available desks, matching meeting rooms, optimizing your commute

    2 of 3: You see it working

    Reading your calendar, finding desks, matching rooms, planning your commute.

  3. Privacy and data screen: what Sync keeps, with controls to keep learning, review, export or forget everything

    3 of 3: You control what it remembers

    See what's stored, manage learning, export your data or erase it all.

The hero flow

The final screens, following Neha through one office day, from a morning that’s already booked to a mid-day re-route she never had to touch.

  1. Home screen: I booked your whole day at 6:31 AM, today’s meetings and the desk, parking and room held

    Step 1

    1 of 6: The day, already booked

    By 6:31 AM Sync has held Neha's desk, parking and room from her calendar.

  2. Recommendation screen suggesting desk D5-08 for the 10 AM

    Step 2

    2 of 6: A suggestion, not a command

    Sync proposes the best-fit space and leaves the final call to Neha.

  3. Why this desk: 92% confidence, the factors it weighed and two alternatives

    Step 3

    3 of 6: Why this desk

    One tap shows exactly why a desk was chosen.

  4. Meeting moved: Design Review moved to 16:30, Atlas is only held until 16:00

    Step 4

    4 of 6: The meeting moves

    At 1:01 PM Design Review shifts to 16:30 and the room is about to run out.

  5. Room moved: Nimbus held for 16:30 and Atlas released, with the reasons

    Step 5

    5 of 6: Re-routed for her

    Sync lets the old room go and holds a new one, with parking.

  6. You’re all set: desk, parking and room confirmed

    Step 6

    6 of 6: Nothing left to do

    Everything is confirmed and Neha never broke focus.

The screens explore auto and confirm-first modes. Their confidence scores need calibration before release; a percentage alone is not evidence that a booking is safe.

Ghost bookings and cancelling, on your terms

Sync’s promise is that it frees up space other people need without ever acting behind your back. These two flows are where that is tested, so they got the most attention.

Catching a ghost booking

A meeting is cancelled but the room is still held in Neha’s name. Sync flags it, says who’s waiting, and hands the decision back to her.

  1. My bookings list with Atlas still held in 40 minutes

    Step 1

    1 of 3: Spotted on the list

    “Atlas · in 40 min” is still held, but plans have changed.

  2. Ghost booking screen: your 3 PM was cancelled, two people are waiting for a Floor 5 room

    Step 2

    2 of 3: Flagged, not deleted

    “Your 3 PM was cancelled.” Sync asks before letting the room go.

  3. Atlas is back in the pool: one hour freed, two people were waiting

    Step 3

    3 of 3: Handed back

    An hour of Floor 5 goes back to the two people waiting.

Ghost-booking cleanup asks before release. Auto rebooking follows the chosen permissions; timed holds follow their visible expiry.

Design decision · HAX: support efficient correction ↗ informed the edit and cancel paths below.

Cancelling and changing by hand

When Neha wants to steer, Sync steps back. Any booking can be opened, changed or cancelled directly, and Sync only re-plans if asked.

  1. Booking detail for desk D5-08 with Change and Cancel buttons

    Step 1

    1 of 3: Open any booking

    Everything that's held, with plain Change and Cancel buttons.

  2. Change booking: change the time, swap the desk, move to another day, or let the agent handle it

    Step 2

    2 of 3: Choose the change

    Change the time, swap the desk, move the day, or let Sync handle it.

  3. Cancel this desk sheet explaining what happens, with Yes, cancel it

    Step 3

    3 of 3: Cancel cleanly

    The space goes straight back to everyone else.

Sync is a default, never a lock-in. A manual override is always one tap away, and “Keep it as it is” is always an option.

Explore the design system and components

To move fast on a big brief, I used Claude Code with the Figma MCP to build Sync’s design system first: colour tokens, type, variables, and components with variants, all inside Figma.

Sync design system cover: Claude Code and Figma connected through MCP.

Typography

Inter

RegularMediumSemi-boldBold

Colour

  • Primary#0066ff
  • Alert#ff8d67
  • Success#4fe3b6

Components

Sync components: Button in each state, Resource Card, Bottom bar, and Input Fields in default, focus and error states

The hard part

How much should an AI do on its own?

Booking a desk is easy. Letting software book it for you is not. The next four sections are about the rules behind Sync: when it acts by itself, when it asks, and how you can review and correct what it did.

How Sync decides

A changing day needs a new plan, not new permissions. This proposed loop separates the agent’s judgement from the rules that constrain its actions.

Design decision · Anthropic: Building effective agents ↗ informed the feedback loop. Flexible planning stays separate from fixed permission and availability checks.

Proposed agent loop · one meeting moves

  1. 01

    Observe

    Calendar change

    Review moves to 16:30. Atlas is unavailable.

  2. 02

    Plan

    Agent proposes

    Find a room that fits the time, team and equipment.

  3. 03

    Check

    Rules constrain

    Check availability, permissions and the user’s autonomy setting.

  4. 04

    Act + verify

    Tool result confirms

    Book only within permission. Report success or ask for help.

Human checkpoint

A fee, another building, an affected colleague or unclear context? Show the change and wait for approval.

Outcome + user correction feed the next plan. Feedback never grants wider permission.

ManualSuggest

You make the booking.

Confirm firstAsk → act

You approve each proposal.

Full autoAct → report

Only inside agreed limits.

Paused means no new actions. The existing plan stays visible; the user can book manually.

Why an agent, and where fixed rules belong

The proposed agent chooses a new plan as the workday changes. Availability checks, permissions and booking confirmation stay deterministic. A fixed “if meeting moves, move room” workflow is enough when the options and steps are predictable.

Calendar content is context, not permission. A production system must restrict connected tools, verify their results and stop for help when it cannot recover.

Reference: Anthropic · workflows and agents ↗

Showing its work

An assistant that works silently feels like a black box. So Sync always shows three things: what it's doing, when it needs you, and what it did.

  1. You can watch it work

    Live steps show progress, such as “Finding a desk near your team”. Cancel stops further work; completed actions stay in the log for review.

  2. It asks before it commits

    When a rule says so, Sync pauses and hands you the decision before anything is booked, and tells you which rule made it ask.

  3. Everything is written down

    The Activity log keeps each action and its reason, with recovery options you can check later.

  1. Booking your Tuesday: each step checked off as Sync works, with a Cancel button

    While it works

    1 of 3: A task you can watch

  2. Reserve these three? A sheet asking to confirm a desk, parking and room

    When it needs you

    2 of 3: It asks before booking

  3. Activity detail for Held Nook 3: why, the rule that allowed it, and Undo

    Afterwards

    3 of 3: A record you can check

Microsoft’s Agent UX principles informed the visible status and Activity record. I explain choices through relevant inputs and outcomes, with details available when needed.

You stay in charge

Trust isn't one feature. It's having a clear answer to the four questions people ask about any assistant.

  • “Why did it pick that?”

    Tap a booking to see the relevant preferences, availability and alternatives. The confidence scores shown in the screens are illustrative, not calibrated.

  • “Can I stop it?”

    One switch pauses Sync. It stops acting on its own, and your bookings stay as they are.

  • “Can I take it back?”

    The concept provides Undo before a booking starts. If a released space is no longer available, recovery must offer alternatives instead of promising an exact reversal.

  • “Can it go beyond what I allowed?”

    The design keeps permission changes explicit: only you can widen its limits. Feedback changes preferences, not what Sync is allowed to do.

  1. Autonomy set to Full auto, with the list of cases where Sync will still ask first

    Limits

    1 of 3: Rules it must follow

  2. You tab: what Sync has learned, each removable, and the Sync is on switch

    Pause

    2 of 3: One switch to stop it

  3. What I did: every action Sync took today, including one that was undone

    Undo

    3 of 3: Review and recover

Design decision · Microsoft HAX: support efficient correction ↗ informed pause, edit and recovery. Undo is offered only when reversal is possible.

Full auto, with limits

Some people want to hand over completely. Full auto is for them. It still reports everything it does, and still asks on the big calls.

  1. 6:31 AM

    Planned overnight

    A lock-screen summary lists what Sync booked while you slept.

  2. 1:01 PM

    The meeting moves, the room follows

    Sync books a new room, lets the old one go, then tells you what changed and why.

  3. Thursday

    A call it won't make alone

    A 12-person workshop only fits in another building, so even Full auto waits for you.

  4. Any time

    One tap back

    Review every action in “What I did”, with Undo where reversal is still possible.

  1. Lock screen notification: Your Tuesday is handled

    1 of 4: Planned overnight

  2. I moved your room: Quarry booked, Atlas released, and why

    2 of 4: Room moved for you

  3. Thursday needs a bigger room: the only fit is in another building, so Sync is asking

    3 of 4: Still asks

  4. Undone: Nook 3 is held for you again

    4 of 4: Undone

Reality check

Does it hold up when things go wrong?

The concept also explores failure, recovery and an interactive prototype. These scenarios still need validation with workplace users.

When things go wrong

The moments that break trust are the ones where something fails. Each one got a clear, helpful screen instead of an error message.

  • No desks are free

    What Sync does: Sync suggests another day or floor. If you delegate the search, it can watch for a match within your limits and ask before making a wider change.

  • A room breaks

    What Sync does: A notification explains the problem, with a replacement already held. Six people are affected, so it waits for your OK.

  • You don't show up

    What Sync does: The hold counts down, then the desk goes back to people who need it.

  • The connection drops

    What Sync does: Home shows your saved plan and what still works. Pending changes need a fresh availability and permission check when the connection returns.

  1. Nothing free on Floor 5 that day, with three alternatives and Let me handle it

    1 of 4: Nothing free

  2. Notification: Your 4 PM lost its room, Sync is holding Atlas

    2 of 4: A room breaks

  3. Check in screen with a hold that expires in 102 minutes

    3 of 4: No-show

  4. Offline home: the saved plan, what still works and what is waiting to sync

    4 of 4: Offline

From Figma to an interactive prototype

I used Claude Code with the Figma MCP to take Sync from screens to an interactive web prototype. Agent behavior and booking data are simulated throughout. It demonstrates the proposed interaction, with no live AI model or booking integrations.

  • Interactive flowsexplore sign-up, booking and autonomy settings
  • Light + darkthemes, checked for readable contrast
  • Simulated datademonstrates proposed behavior; no live integrations
Try the live prototype at sync-web-phi.vercel.app (opens in a new tab)
  1. Book a space in the light theme: free desks ranked best for you first

    Light

    1 of 4: Browse

  2. Floor 5 map in the dark theme with your desk and your team highlighted

    Dark

    2 of 4: Live map

  3. Notifications in the light theme: booked, moving, an AV fault and an approved change

    Light

    3 of 4: Notifications

  4. What do you need? sheet in the dark theme: a desk, a meeting room, a parking bay, or just tell me

    Dark

    4 of 4: Book anything

What I learned

This sprint produced a design concept and a simulated prototype. Desk research shaped the scenarios; real-world reliability and usability remain to be validated.

  1. Design the permission, then the screen.

    Designing an agent shifted my focus from the steps a person takes to the actions software may take for them. I made autonomy explicit for desks, rooms and parking, with exceptions that still ask.

  2. Background work needs a visible record.

    Microsoft’s Agent UX principles helped me treat status, reasons and history as part of the main flow. “Why this desk” sits beside the recommendation; Activity keeps decisions available for review.

  3. Recovery is more than an Undo button.

    HAX’s correction guidance pushed me to design edit, pause and manual takeover. A released room can be taken by someone else, so recovery must offer another option when exact reversal is impossible.

  4. A prototype makes the next questions clearer.

    Anthropic’s distinction between workflows and agents helped separate flexible planning from fixed booking checks. The simulated prototype explores that interaction; the next step is testing comprehension, recovery and permission boundaries with workplace users.

Next validation · questions to test

Understand

Can someone predict when Sync will ask? Measure correct predictions.

Recover

Can they fix a wrong booking? Measure completion, time and failed recovery.

Respect limits

Test stale data, lost access and partial booking failures. Count unauthorized actions.

Sync

Aims to ease the daily booking ritual, and return the space nobody is using.

Sync Agent: two Sync screens, the welcome screen and the Home dashboard with the day booked. Callouts: Desk H2-12 near Design, booked 3:30 PM, before you asked; Parking P12, Level B1, you drive on Tuesdays.
Responsibilities
Desk research · UX/UI design · Prototyping
Timeline
1-day sprint
Industry
Workplace technology

Project overview

A booking assistant for hybrid offices: plan the desk, room and parking around a changing workday. I designed when it acts, when it asks, and how people stay in control. Agent behavior and booking data are simulated in the prototype.

Three Sync screens: Atlas is back in the pool, the Home dashboard with the day booked, and Booking your Tuesday

The context

Modern corporate offices increasingly run on a hot-desking model, where employees no longer own a fixed seat. Hybrid workers come in only on selected days, mainly for meetings and collaboration, and must book whatever resources they need each time.

The challenge

Hot-desking freed the floor plan but pushed the work onto people. Every day meant re-booking a desk, chasing rooms that vanished when a meeting moved, and living with ghost bookings that hoarded space nobody used.

  1. A tired employee juggling a calendar, a desk map and a parking list on floating screens

    Manual, repetitive booking every single morning

  2. A meeting slot turning red and moving on a floating schedule inside a glass meeting room

    Rooms disappearing the moment a meeting shifts

  3. An empty office desk still marked as booked, with faint outlines of people who never came

    Ghost bookings holding space others are waiting on

Brief

Design an agentic AI that handles all of this proactively, while keeping the user firmly in control: actions explained, correction available, and autonomy bounded by the limits the user sets.

My approach

Before shaping the interactions, I reviewed Microsoft’s Agent UX principles, its HAX guidance on correction, and Anthropic’s Building effective agents. I used them to define three decisions: make background work visible, keep correction close to the action, and put explicit checkpoints around autonomy.

A booking form asks people to choose every resource. An agent asks them to delegate a goal, then needs to show what it may do, what changed, and when it needs help. I designed those boundaries before polishing the screens.

The guidance behind the design decisions

01

Make background work visible

Microsoft connects trust with transparency and control. I translated that into live task steps, “Why this desk” and an Activity record.

UX design for agents ↗
02

Make correction part of the flow

HAX recommends making AI mistakes easy to correct. I kept edit, pause and manual takeover close to each action, with alternatives when Undo cannot restore a space.

HAX · Support efficient correction ↗
03

Separate planning from permission

Anthropic describes agents checking tool results and returning to people at checkpoints. I applied that to a plan → check → act loop, with confirm-first and exceptions in Full auto.

AnthropicBuilding effective agents ↗

These are my applications of the guidance to Sync. The sources informed the concept; they do not validate its usability or reliability.

My approach, as a FigJam board. Understanding the Context and user: What is Hot-desking? How is it currently managed? Who are the target users? How their day looks like? Identifying friction points: What is the friction? When does it occur? How does it affect the users? Ideating on Agent Interaction: What actions can the agent perform? How much control will the agent have? When does the Agent Act? These lead to Connecting the Business Needs and user goals: Efficient usage of resources = Efficient usage of time. Which leads to Designing the Solution: Scalable Design system, Human in the loop, Context aware agent, Privacy and security control, Handy Manual override.

Design strategy

Business Objective

Manage and utilise meeting rooms, desks and parking efficiently, reducing the friction of booking and ensuring a seamless hybrid work culture.

Solution Overview

An agentic AI-powered app that works quietly in the background, reducing the load on employees.

  1. Remove repetitive manual booking effort
  2. Free up "ghost-booked" resources for others
  3. Maximise utilisation of shared workspace resources
  • Reads contextIntegrates with calendar, location & past booking preferences.
  • Re-routes automaticallyProposes or rebooks when plans change, within your permissions.
  • Books proactivelyChalks out office days and auto-books the resources you'll need.

Who is Sync for?

Concept personas based on desk research, not interview participants. Their quotes express the needs explored in this concept.

Portrait of Neha

Neha

Product Manager

Demographics

Age:
28
Work Style:
Highly collaborative, constantly switching between focus work and meetings.
About:
Curious and adaptable product manager who enjoys solving complex problems and working closely with different teams.

"My day changes by the hour, so my bookings should keep up without me babysitting them."

Pain Points

  • Constant manual updates to existing bookings
  • Manual filtering for the right room adds cognitive load
  • Different resources for each meeting

Needs and Goals

  • Meeting rooms with video-conferencing support
  • To sit near different teams depending on the day's focus
  • Context-specific spaces, e.g. a room with a large whiteboard
Portrait of Kabir

Kabir

Senior Data Analyst

Demographics

Age:
29
Work Style:
Deep-focus worker, in office only 1 to 2 days a week, prefers a quiet environment.
About:
A methodical senior data analyst who does his best thinking in long, uninterrupted blocks and plans his office days around the setup he needs.

"I drive in for deep-focus days, and if my setup and parking aren't guaranteed, the trip is wasted."

Pain Points

  • Struggles to find parking spots
  • Misses his equipment setup because he forgot to pre-book 48 hrs in advance
  • Doesn’t always get a quiet space

Needs and Goals

  • Specific equipment such as a dual-monitor setup.
  • Prefers to sit in quiet zones for concentration.
  • A guaranteed parking spot, since he commutes a long distance.

User journey map

A desk-research journey hypothesis, ready to test with workplace users.

User journey map: four stages of a booking day
Plan the dayBook workspaceBook supporting needsChanges happen
User ActionsChecks meetings, office location and timingsOpens booking platform and searches for a deskSeparately books parking and meeting roomsManually modifies/ cancels bookings
EmotionsNeutralFrustratedOverwhelmedAnnoyed
Pain PointsInformation is spread across calendar and workplace systemsManual search and selectionMultiple systems and separate bookingsChanges require repeated manual work
OpportunitiesUnderstand the user's entire workdayRecommend the right workspaceBundle desk + room + parkingAutomatically adapt bookings to changes

Ideas before interfaces

Ideation notes: brainstormed how agentic AI can use background data and integrate itself, fetching the calendar and past habits, to eliminate manual booking; explored how agentic AI browsers like ChatGPT Atlas and Comet work in real time to understand agentic AI capabilities. Quick concept sketching: Concept A, a summary card (“AI booked desk near design team 2:00 PM” with Change and Keep); Concept B, a calendar edit (Today, 9:00 AM, 10:00 AM Team Sync edited, as agent rerouted, 11:00 AM).Two home screen sketches: one where the meeting was delayed to 3 PM and the agent released the old room and booked Room C on the 3rd floor, with Undo and Edit, showing the AI doing the work in the background, automatically syncing and rescheduling; and one reading “Good morning Neha, your Tuesday AI setup: desk near tech team, board room C at 3 PM, edit preferences”, noted as showing transparent controls, with a + icon added to still allow manual booking in edge cases.

Meet Sync

A concept for an assistant that plans your spaces, explains its choices, and acts within the permissions you set.

  1. Eliminates mental load

    When plans change, Sync proposes or handles rebooking within your chosen limits.

  2. Reduces ghost bookings

    Forgotten bookings are flagged the moment a meeting is delayed or cancelled, and released once you agree, freeing them for everyone else.

  3. Keeps the human in control

    Actions stay visible, with controls to undo, edit or take over.

Agent interactions

Design decision · Following Microsoft’s transparency and control principles ↗, I gave autonomy, status and memory their own visible controls.

Three screens set up the relationship with the assistant: how much it may do on its own, what it's doing right now, and what it remembers about you.

  1. Autonomy level screen: How much should I decide alone? Full auto, Confirm first (selected) and Manual

    1 of 3: User-defined autonomy

    Users decide how much Sync can handle on its own. From fully automatic booking to confirmation-first, autonomy can be adjusted per resource.

  2. Agent screen: Lining up your week, reading calendar, finding available desks, matching meeting rooms, optimizing your commute

    2 of 3: Transparent agent actions

    Sync shows what it is doing while it works, reading the calendar, finding desks, matching rooms, and optimizing the commute, so the agent never feels like a black box.

  3. Privacy and data screen: what Sync keeps, with controls to keep learning, review, export or forget everything

    3 of 3: Control over Sync Memory

    Sync makes its data boundaries explicit. Users can see what is stored, manage learning, export their data, or erase everything at any time.

The hero flow

The final high-fidelity screens, following Neha through a single office day, from an auto-booked morning to a mid-day reroute she never had to touch.

  1. Home screen: I booked your whole day at 6:31 AM, today’s meetings and the desk, parking and room held

    Step 1

    1 of 6: The day, already booked

    By 6:31 AM the agent has held Neha's desk, parking and room straight from her calendar.

  2. Recommendation screen suggesting desk D5-08 for the 10 AM

    Step 2

    2 of 6: A recommendation, not a command

    Sync proposes the best-fit space and always leaves the final call to Neha.

  3. Why this desk: 92% confidence, the factors it weighed and two alternatives

    Step 3

    3 of 6: Transparent reasoning

    One tap reveals exactly why a desk was chosen, with no black box.

  4. Meeting moved: Design Review moved to 16:30, Atlas is only held until 16:00

    Step 4

    4 of 6: The meeting moves

    At 1:01 PM Design Review shifts to 16:30 and the held room is about to expire.

  5. Room moved: Nimbus held for 16:30 and Atlas released, with the reasons

    Step 5

    5 of 6: Re-routed automatically

    The agent releases the old room and secures a new one, with parking, on its own.

  6. You’re all set: desk, parking and room confirmed

    Step 6

    6 of 6: Nothing left to do

    Everything is confirmed and Neha never once broke focus.

The screens explore auto and confirm-first modes. Their confidence scores need calibration before release; a percentage alone is not evidence that a booking is safe.

Ghost bookings and cancelling, on the user's terms

The whole promise of Sync is that it frees space others need without ever acting behind the user's back. These two flows are where that principle is tested, so they got the most attention.

The ghost-booking guardrail

A meeting is cancelled but the room stays held in Neha's name. Sync flags it as a ghost booking, explains who's waiting, and hands the decision back to her.

  1. My bookings list with Atlas still held in 40 minutes

    Step 1

    1 of 3: Spotted on the list

    "Atlas · in 40 min" is still held while plans have quietly changed.

  2. Ghost booking screen: your 3 PM was cancelled, two people are waiting for a Floor 5 room

    Step 2

    2 of 3: Flagged, not deleted

    "Your 3 PM was cancelled." The room is still held. Sync asks before releasing.

  3. Atlas is back in the pool: one hour freed, two people were waiting

    Step 3

    3 of 3: Handed back

    One hour of Floor 5 returns to the pool: 2 people were waiting, 0 ghosts left.

Ghost-booking cleanup asks before release. Auto rebooking follows the chosen permissions; timed holds follow their visible expiry.

Cancelling and changing manually

When Neha wants to steer, the agent steps back. Any booking can be opened, changed, or cancelled directly, with the AI offering to re-plan only if asked.

  1. Booking detail for desk D5-08 with Change and Cancel buttons

    Step 1

    1 of 3: Open any booking

    Full detail of what's held, with plain "Change" and "Cancel" controls.

  2. Change booking: change the time, swap the desk, move to another day, or let the agent handle it

    Step 2

    2 of 3: Choose the change

    Change the time, swap the desk, move to another day, or let the agent handle it.

  3. Cancel this desk sheet explaining what happens, with Yes, cancel it

    Step 3

    3 of 3: Cancel cleanly

    Confirm the cancel and the space is freed for the pool immediately.

“The agent is a default, never a lock-in. Manual override is always one tap away, and "Keep it as it is" is always an option.”

Explore the design system and components

To move fast on a big brief, I used Claude Code with the Figma MCP to build Sync's design system first: colour tokens, type, variables, and components with variants, all inside Figma.

Sync, your AI agent for a seamless workplace experience: Design System. Claude Code and Figma connected through MCP.

Typography

Inter

RegularMediumSemi-BoldBold

Color Palette

  • #0066ff
  • #ff8d67
  • #4fe3b6

Components

Sync components on a dark canvas: Button in each state, Resource Card, Bottom bar, and Input Fields in default, focus and error states

The hard part

How much should an AI do on its own?

Booking a desk is easy. Letting software book it for you is not. The next four sections are about the rules behind Sync: when it acts by itself, when it asks, and how you can review and correct what it did.

How Sync decides

A changing day needs a new plan, not new permissions. This proposed loop separates the agent’s judgement from the rules that constrain its actions.

Design decision · Anthropic: Building effective agents ↗ informed the feedback loop. Flexible planning stays separate from fixed permission and availability checks.

Proposed agent loop · one meeting moves

  1. 01

    Observe

    Calendar change

    Review moves to 16:30. Atlas is unavailable.

  2. 02

    Plan

    Agent proposes

    Find a room that fits the time, team and equipment.

  3. 03

    Check

    Rules constrain

    Check availability, permissions and the user’s autonomy setting.

  4. 04

    Act + verify

    Tool result confirms

    Book only within permission. Report success or ask for help.

Human checkpoint

A fee, another building, an affected colleague or unclear context? Show the change and wait for approval.

Outcome + user correction feed the next plan. Feedback never grants wider permission.

ManualSuggest

You make the booking.

Confirm firstAsk → act

You approve each proposal.

Full autoAct → report

Only inside agreed limits.

Paused means no new actions. The existing plan stays visible; the user can book manually.

Why an agent, and where fixed rules belong

The proposed agent chooses a new plan as the workday changes. Availability checks, permissions and booking confirmation stay deterministic. A fixed “if meeting moves, move room” workflow is enough when the options and steps are predictable.

Calendar content is context, not permission. A production system must restrict connected tools, verify their results and stop for help when it cannot recover.

Reference: Anthropic · workflows and agents ↗

Showing its work

An assistant that works silently feels like a black box. So Sync always shows three things: what it's doing, when it needs you, and what it did.

  1. You can watch it work

    Live steps show progress, such as “Finding a desk near your team”. Cancel stops further work; completed actions stay in the log for review.

  2. It asks before it commits

    When a rule says so, Sync pauses and hands you the decision before anything is booked, and tells you which rule made it ask.

  3. Everything is written down

    The Activity log keeps each action and its reason, with recovery options you can check later.

  1. Booking your Tuesday: each step checked off as Sync works, with a Cancel button

    While it works

    1 of 3: A task you can watch

    Each step appears as it happens, with Cancel always there.

  2. Reserve these three? A sheet asking to confirm a desk, parking and room

    When it needs you

    2 of 3: It asks before booking

    Nothing is held until you confirm.

  3. Activity detail for Held Nook 3: why, the rule that allowed it, and Undo

    Afterwards

    3 of 3: A record you can check

    Why it acted, what allowed it, and an Undo.

Microsoft’s Agent UX principles informed the visible status and Activity record. I explain choices through relevant inputs and outcomes, with details available when needed.

You stay in charge

Trust isn't one feature. It's having a clear answer to the four questions people ask about any assistant.

  • “Why did it pick that?”

    Tap a booking to see the relevant preferences, availability and alternatives. The confidence scores shown in the screens are illustrative, not calibrated.

  • “Can I stop it?”

    One switch pauses Sync. It stops acting on its own, and your bookings stay as they are.

  • “Can I take it back?”

    The concept provides Undo before a booking starts. If a released space is no longer available, recovery must offer alternatives instead of promising an exact reversal.

  • “Can it go beyond what I allowed?”

    The design keeps permission changes explicit: only you can widen its limits. Feedback changes preferences, not what Sync is allowed to do.

  1. Autonomy set to Full auto, with the list of cases where Sync will still ask first

    Limits

    1 of 3: Rules it must follow

    Review the situations that require approval, even in Full auto.

  2. You tab: what Sync has learned, each removable, and the Sync is on switch

    Pause

    2 of 3: One switch to stop it

    Sync stops acting on its own. Your bookings stay.

  3. What I did: every action Sync took today, including one that was undone

    Undo

    3 of 3: Review and recover

    What it did today, why, and a way back.

Design decision · Microsoft HAX: support efficient correction ↗ informed pause, edit and recovery. Undo is offered only when reversal is possible.

Full auto, with limits

Some people want to hand over completely. Full auto is for them. It still reports everything it does, and still asks on the big calls.

  1. 6:31 AM

    Planned overnight

    A lock-screen summary lists what Sync booked while you slept.

  2. 1:01 PM

    The meeting moves, the room follows

    Sync books a new room, lets the old one go, then tells you what changed and why.

  3. Thursday

    A call it won't make alone

    A 12-person workshop only fits in another building, so even Full auto waits for you.

  4. Any time

    One tap back

    Review every action in “What I did”, with Undo where reversal is still possible.

  1. Lock screen notification: Your Tuesday is handled

    1 of 4: Planned overnight

    The day is planned before you wake.

  2. I moved your room: Quarry booked, Atlas released, and why

    2 of 4: Room moved for you

    A new room, with the reasons.

  3. Thursday needs a bigger room: the only fit is in another building, so Sync is asking

    3 of 4: Still asks

    Another building, so it waits for you.

  4. Undone: Nook 3 is held for you again

    4 of 4: Undone

    The previous space is restored in this example.

Reality check

Does it hold up when things go wrong?

The concept also explores failure, recovery and an interactive prototype. These scenarios still need validation with workplace users.

When things go wrong

The moments that break trust are the ones where something fails. Each one got a clear, helpful screen instead of an error message.

  • No desks are free

    What Sync does: Sync suggests another day or floor. If you delegate the search, it can watch for a match within your limits and ask before making a wider change.

  • A room breaks

    What Sync does: A notification explains the problem, with a replacement already held. Six people are affected, so it waits for your OK.

  • You don't show up

    What Sync does: The hold counts down, then the desk goes back to people who need it.

  • The connection drops

    What Sync does: Home shows your saved plan and what still works. Pending changes need a fresh availability and permission check when the connection returns.

  1. Nothing free on Floor 5 that day, with three alternatives and Let me handle it

    1 of 4: Nothing free

    Three ways round it, or let Sync handle it.

  2. Notification: Your 4 PM lost its room, Sync is holding Atlas

    2 of 4: A room breaks

    A notification with the fix already held.

  3. Check in screen with a hold that expires in 102 minutes

    3 of 4: No-show

    The hold counts down, then the desk goes back.

  4. Offline home: the saved plan, what still works and what is waiting to sync

    4 of 4: Offline

    The saved plan, what still works, what is waiting.

From Figma to an interactive prototype

I used Claude Code with the Figma MCP to take Sync from screens to an interactive web prototype. Agent behavior and booking data are simulated throughout. It demonstrates the proposed interaction, with no live AI model or booking integrations.

  • Interactive flowsexplore sign-up, booking and autonomy settings
  • Light + darkthemes, checked for readable contrast
  • Simulated datademonstrates proposed behavior; no live integrations
Try the live prototype at sync-web-phi.vercel.app (opens in a new tab)
  1. Book a space in the light theme: free desks ranked best for you first

    Light

    1 of 4: Browse

    Every free desk, best for you first.

  2. Floor 5 map in the dark theme with your desk and your team highlighted

    Dark

    2 of 4: Live map

    Who sits where, and your desk on the floor.

  3. Notifications in the light theme: booked, moving, an AV fault and an approved change

    Light

    3 of 4: Notifications

    Each one opens the flow it is about.

  4. What do you need? sheet in the dark theme: a desk, a meeting room, a parking bay, or just tell me

    Dark

    4 of 4: Book anything

    Desk, room or parking, from any screen.

What I learned

This sprint produced a design concept and a simulated prototype. Desk research shaped the scenarios; real-world reliability and usability remain to be validated.

  1. Design the permission, then the screen.

    Designing an agent shifted my focus from the steps a person takes to the actions software may take for them. I made autonomy explicit for desks, rooms and parking, with exceptions that still ask.

  2. Background work needs a visible record.

    Microsoft’s Agent UX principles helped me treat status, reasons and history as part of the main flow. “Why this desk” sits beside the recommendation; Activity keeps decisions available for review.

  3. Recovery is more than an Undo button.

    HAX’s correction guidance pushed me to design edit, pause and manual takeover. A released room can be taken by someone else, so recovery must offer another option when exact reversal is impossible.

  4. A prototype makes the next questions clearer.

    Anthropic’s distinction between workflows and agents helped separate flexible planning from fixed booking checks. The simulated prototype explores that interaction; the next step is testing comprehension, recovery and permission boundaries with workplace users.

Next validation · questions to test

Understand

Can someone predict when Sync will ask? Measure correct predictions.

Recover

Can they fix a wrong booking? Measure completion, time and failed recovery.

Respect limits

Test stale data, lost access and partial booking failures. Count unauthorized actions.