From task brief to finished result

Five days of product decisions

I redesigned an AI agent workspace with the founder and engineering lead, moving from a crowded dashboard to a simpler Playground with detail on demand.¹

This case records the decisions and tradeoffs behind the home and main task workflow. Supporting screens, reviews and handoff work continued after the five-day sprint.

A brand manager’s first task

Switch lets companies give tasks and files to AI agents and receive imagery, documents or presentations. We were designing for brand managers with little AI experience. They were trying the agent before briefing an agency and needed a result they could use.

The interface had duplicate agent lists, scattered activity and results that were hard to find. I owned the redesign, working with the founder and engineering lead. We spent five working days on the home and main workflow, then continued with supporting screens and reviews.

I drew on the platform walkthrough, engineering documentation, six meeting records from 14–22 October and the team’s account of customer usage. We had no new usability study or post-launch measurements.

01 / A simpler Playground, with detail on demand

At kickoff we considered separate work and control modes. At Checkpoint 1, we chose one home with relevant agents, ongoing work and completed outputs.

On 16 October, we still preferred a richer view. But runs, insights and artifacts competed below the search bar. Putting them in separate panels helped organise the page; we still needed to decide what a new user should see first.

At the 19 October review, we narrowed the home to two jobs: start work with an agent and retrieve the deliverable. We reduced recent runs to two or three compact items. By 20 October, we had chosen a simple day-one Playground with an optional advanced sidebar for experienced users.

Inside a task, Settings, Preview and Insights give users access to supporting information while they work.

Settings, Preview and Insights on demand

Close

1 / 3

Moving detail into the sidebar added a step for anyone checking status. We kept requests needing attention visible and added a persistent activity signal.

We moved reports and connections toward administration and settings.

Discover agents by what they produce

Close

1 / 1

I separated the reusable agent setup from each task and its deliverables. At the later review, we chose “Task” for a run: using an agent with new inputs creates a task.

The engineering documentation described personal conversations and shared files. We could show colleagues’ deliverables from Library. Showing their live tasks would need new backend capability; a navigation change could not grant that access.

02 / Keep background work within reach

On 19 October, we chose a small persistent indicator that would open task details. Users could check background work while starting another task or finding an output.

Live activity shows running tasks. Notifications cover alerts, received shares and connection requests. Combining them could bury a request among processing updates. We wanted access to both while browsing Playground, Agents or Library.

Open the monitor to inspect background work

Close

1 / 1

The recent export shows the monitor open, with counts and links for work needing attention. It has a separate notification bell. I have not verified the proposed two-tab interaction or its production implementation.

The platform supported live events and Working, Needs you, Review, Done, Stopped and Failed states. I used these to distinguish active work from a pause and explain what the user needed to do next.

Retrieve shared results

Close

1 / 1

03 / Ask enough to get the brief right

My early aim was to minimise questions before generation. Engineering pushed back: missing details could mean generating the image again. I changed the goal to asking enough to produce a useful result, while reusing known facts and showing defaults.

I separated input checks before a model call from questions about orientation, palette or target medium. We accepted some preparation effort to avoid rework. We still needed to confirm support for recognising the wrong product in an uploaded image.

The engineering analysis reported follow-up corrections in a small customer sample, mainly about art direction. I put conversation and result in one workspace and removed the intermediate popup from the proposed flow. Users could request a change or download a good first result.

Prepare the brief, then review the result

Close

1 / 2

We dropped the fixed three-variant requirement to prioritise quality. I kept download resizing separate from generating a new composition. Image follow-up regenerated the output, so the controls had to avoid promising a precise edit or version restoration.

Download the formats selected during setup

Close

1 / 1

Scope and remaining decisions

We deferred the node-based workflow visualizer on 19 October to keep time for the main delivery flow.

Customers had asked for human approval. The platform supported approval of a paused tool call by the person who started it; colleague artwork sign-off needed further work. The prototype’s approval control remained a proposal.

At the 22 October review, the founder proposed applying each customer’s brand colours. I raised the difficulty of maintaining contrast and proposed a neutral base with selected brand accents. We had not settled the compromise or measured its effect.

Login and invited-user onboarding

Close

1 / 2

Delivery and next validation

I used the existing tokens and components, kept a capability record and built the prototype with Claude. I made the design decisions and checked generated copy against platform constraints.

The five-day sprint established the home and main workflow. At the 20 and 22 October reviews, we still had supporting UI, onboarding, sharing and handoff work to finish. Production rollout and customer outcomes remain unconfirmed.

Next, I would test whether a new brand manager can complete a task with the advanced sidebar closed, notice a request for input, refine the result and retrieve it later.

Close