The workspace got too clever. Here is what I removed.

The first version of the workspace composer tried to expose too much of the internal model. There were top-level directions like Ask, Backtest, Research, and Swarm. Then there was another Task versus Discuss toggle. Then a credit estimate that often showed the same number. None of this helped the user think.

The problem was not visual polish

You can make a bad control beautiful. That does not make it useful. The real issue was cognitive load: two layers of mode selection created eight possible states, while the backend was mostly choosing different skills behind the scenes.

What changed

• Kept the first layer: Ask, Backtest, Research, Swarm.
• Removed the Task / Discuss toggle because it implied a billing contract that did not really exist.
• Removed static credit estimates from the composer. Real cost belongs after execution, not as a guessed badge.
• Marked Swarm as an advanced mode because parallel agents can spend more tokens and should not feel like a casual default.

The session list needed the same honesty

The sidebar used to show "active" without saying what active meant. That is not a state, it is a shrug. It now shows whether the session is open or when it was updated, and session actions live where users expect them: rename and delete in a small context menu.

“Less UI is not minimalism. Less UI is fewer wrong promises.”

The rule going forward

If a control maps to a real backend contract, it can stay. If it is only exposing an internal implementation detail, it needs to earn its place. The workspace should feel like asking a sharper question, not operating a control panel.