Summary
Offering Claude and Codex in the agent dropdown is currently ambiguous: users can't tell which model provider or subscription the harness runs against. This has been strong, recurring feedback, and I couldn't get a single user to understand the value prop and how it actually works without a lengthy explainer.
Main blocker: Which provider/subscription is being used?
Most users assume that picking Claude runs against their own Claude Code subscription, and they'd love to keep using it inside our Agents UI. In reality the harness routes through the Copilot proxy and counts against their GitHub Copilot tokens.
Impact: Users don't realize this isn't their existing Claude Code subscription and instead consumes Copilot tokens — a serious trust risk once they find out.
Model picker vs. agent picker
Developers use the same umbrella terms ("Claude", "Codex") for both model families and the harness products. The model picker and the agent picker don't differentiate their intent enough, so it's unclear what a selection actually means.
Impact: Users don't understand the agent picker and avoid using it.
Prior art
The Claude tooltip in the editor window is a nice pattern worth building on.
Ideas (not committed)
- Improve the labels and hover/tooltip details in the Agents window to explain the provider/subscription and billing.
- Extend the editor-window Claude tooltip to the Codex harness.
- Add equivalent tooltips in the Agents window.
- A one-off explainer the first time a user picks a 3rd-party agent.
- Make the active subscription explicit, with an easy toggle (e.g. disabling the Copilot proxy for Claude — foundation already started).
/cc @TylerLeonhardt @Giuspepe
Summary
Offering Claude and Codex in the agent dropdown is currently ambiguous: users can't tell which model provider or subscription the harness runs against. This has been strong, recurring feedback, and I couldn't get a single user to understand the value prop and how it actually works without a lengthy explainer.
Main blocker: Which provider/subscription is being used?
Most users assume that picking Claude runs against their own Claude Code subscription, and they'd love to keep using it inside our Agents UI. In reality the harness routes through the Copilot proxy and counts against their GitHub Copilot tokens.
Impact: Users don't realize this isn't their existing Claude Code subscription and instead consumes Copilot tokens — a serious trust risk once they find out.
Model picker vs. agent picker
Developers use the same umbrella terms ("Claude", "Codex") for both model families and the harness products. The model picker and the agent picker don't differentiate their intent enough, so it's unclear what a selection actually means.
Impact: Users don't understand the agent picker and avoid using it.
Prior art
The Claude tooltip in the editor window is a nice pattern worth building on.
Ideas (not committed)
/cc @TylerLeonhardt @Giuspepe