I’ve been thinking a lot about what the best interaction UI with agents should be at the current stage.
One approach is something like the Codex app. But the problem is that the data is private/local. In the future, if we want to sync chat/task history across multiple devices and keep reliable backups, this may become inconvenient.
I feel the truly valuable form might be a Web UI. I’ve always been curious: why do we have to build a Codex app, instead of building a sister product similar to chatgpt — for example, something like chatagent?
If ChatGPT had a feature to connect to a local machine via SSH, and we could complete agent work directly inside ChatGPT, wouldn’t that be a great interaction model? It could support multi-task management and long-running task management as well.
I think terminal UIs (TUI) and IDE integrations are transitional artifacts.
I wrote some articles to express these views:
I’ve written about this view in more detail elsewhere, but the core idea is:
Right now, a lot of “agent programming UIs” (especially CLI/TUI-style tools) feel like a temporary phase rather than an end state.
In the short term, TUI makes sense because it’s lightweight, scriptable, and easy to ship. But I suspect that a few months down the road, many developers will drift back toward GUI-first experiences and eventually IDE integration again.
That said, today I personally don’t want to live inside an IDE. I don’t want to stare at every line of code. What I care about most is being able to clearly answer:
- What is the agent doing right now?
- Why is it doing it?
- What assumptions is it making, and what state is it relying on?
So the UI shouldn’t optimize for “manual editing speed” first — it should optimize for observability, traceability, and task-level control. If we can make those primitives first-class (state, rationale, progress, branching, rollbacks, approvals), then the “right” surface might naturally look more like a web-based task/thread system than a traditional terminal or IDE.
Another example is the Discourse forum we’re using right now — it’s already a near-perfect place for agent interaction: team collaboration, and multi-task management becomes simple, like in the screenshot below:
