AI agentのUIは、チャット欄ではなくレビュー面になるThis article argues that the core UI for AI agents should shift from chat…

The question of how people should interact with AI agents is becoming a practical design problem rather than a research curiosity. A recent blog post argues that the default choice—putting a chat box at the center of an agent's interface—works well for demos but poorly for real work, and that the primary surface should instead become something closer to a review panel.

The reasoning starts with a familiar temptation. When building an AI agent's UI, the instinct is to place a chat field first: the user asks in natural language, the agent responds, and the user follows up with more questions. This pattern feels distinctly "AI-like" and tends to look impressive in demonstrations. Conversational interfaces have become the default in part because tools like ChatGPT made them the popular mental model for interacting with language models.

But the author contends that the hard part of using an agent at work is not the moment of conversation. The friction shows up afterward, when a person needs to understand what the agent actually did: which sources it read, which tools it invoked, which diffs it produced, and which decisions it handed back to a human. If those steps are not legible, the post argues, what remains at the end is a long transcript and a set of changes that are difficult to review.

From this, the article proposes that the center of agent UI will gradually move from the chat field to what it calls a review surface. The idea is to treat the interface less as a place to talk with an AI and more as a place that decomposes the agent's work into units a human can inspect and take responsibility for. Rather than scrolling a linear conversation, a reviewer would see the agent's actions and judgments broken into discrete, acceptable pieces.

This framing sits within a broader industry conversation about human-in-the-loop design, where a system defers certain decisions to people at defined checkpoints rather than acting autonomously end to end. The concept is especially relevant for agents that call external tools and make changes to real systems, because each tool call and each modification carries consequences that someone may need to approve, reject, or amend.

The argument also echoes patterns already visible in agentic coding tools. Products such as GitHub Copilot, Cursor, and other assistants have increasingly leaned on diff views and pull-request-style workflows, where the model's proposed changes are shown as reviewable deltas rather than buried in prose. Code review is a mature discipline with established conventions—inline comments, side-by-side diffs, approval gates—and it offers a ready template for how agent output might be presented for human oversight. Seen this way, the post's review surface is an attempt to generalize that pattern beyond code to any domain where an agent produces changes.

Several practical implications follow from the author's position, though the post presents them as a direction rather than a finished specification. Making an agent's work reviewable likely requires surfacing its tool calls, the inputs and outputs of those calls, and the reasoning or evidence behind a given diff. It also implies clear handoff points, where the agent explicitly returns control to a human for a decision instead of continuing silently. A transcript can still exist as an audit trail, but under this view it becomes a secondary artifact rather than the primary interface.

None of this necessarily means chat disappears. Natural language remains a convenient way to issue instructions and ask clarifying questions, and the post appears to treat conversation as one component rather than the whole. The claim is narrower and more about emphasis: as agents take on more consequential tasks, the interface's center of gravity is likely to shift toward whatever helps a person verify and accept the work.

For teams building agent products, the takeaway is a design prompt worth weighing. A chat-first layout is easy to ship and easy to demo, but it may push the cost of review onto users at the worst possible moment. Designing for legible, unit-by-unit review from the start could make agents more trustworthy in production, even if it is less immediately eye-catching than a blinking cursor in a chat box.