Anthropic shipped Claude Routines in April 2026. The feature runs multi-step workflows in the background - no prompt required - and when a run completes, each one creates a reviewable Claude Code session where users can see what the agent did after the fact, as a trace timeline rather than a conversation. That design choice is easy to miss. Anthropic didn't put the output back in the chat window. They gave it a different surface entirely. That's the tell.
The consensus position right now is that chat is a fine interface for AI agents - familiar, low-friction, already where people spend their day in Slack or Teams. I think that's wrong. Not because chat is bad, but because it was designed for one thing and is now being pressed into service for something fundamentally different. The cases where it breaks are becoming the normal cases.
What chat was actually built for
The conversational frame is so dominant that "talking to the AI" has become synonymous with "using the AI." The metaphor made sense when the interaction model was prompt-response: you ask, it answers, you evaluate. One turn. Maybe ten. That model fits exploration - when you don't know what you want, when the task is drafting or brainstorming or summarising. Chat works when the user doesn't yet know what they want. It fails when the user knows exactly what they want, and the blank text box becomes a tax on every interaction.
The trouble is that agents aren't exploration tools. They're execution tools. And execution has different interface requirements: state visibility, approval queues, audit logs, the ability to intervene mid-task. None of those exist in a scrolling message thread.
A chat interface is a good input primitive. It is a bad operating environment.
Where the chat model structurally fails
Consider three things a team might want an AI agent to do. Monitor a deployment pipeline and alert on anomalies. Flag a support ticket that's escalating before a human reads it. Notice when a nightly batch job produces out-of-range results. In each case, the chat model introduces a structural bottleneck: the work cannot happen until someone opens the interface and asks. The value of the AI is time-sensitive - it matters when the anomaly is caught, not whether it's caught eventually. Chat defeats the purpose.
There's also a concurrency problem. A chat interface creates a single conversation thread. Real workflows involve parallel tasks, multiple event streams, and concurrent processes. You cannot delegate ten things to a chat interface the way you would to a capable colleague.
The failure mode at the product level is predictable. The team ships a chat UI for a monitoring or automation task, and nobody uses it. Users don't remember to ask. Or they ask once, get a useful response, and never return. The AI creates value only when actively queried - which means it creates value rarely. The product metrics look bad, the team concludes that "AI doesn't work here," and the feature gets cut. The model gets blamed for an interface problem.
The underlying architecture question is the same. Where a chatbot listens for a user message, an ambient agent listens for a system event - a file change, a database write, a webhook, a scheduled trigger. Where a chatbot produces a response that the user reads, an ambient agent may produce an action that updates a record, sends a notification, or triggers another workflow. The human-in-the-loop is still present, but the interface is audit logs and approval flows, not a chat box.
The steelman: chat as the right default
The case for chat is real and worth taking seriously. Chat is where work already happens. Slack and Teams are not going away. The cognitive cost of context-switching to a separate operations dashboard is non-trivial, and most enterprise teams are already over-tooled. If an agent can surface in the same thread where a decision is being made, that's often better than routing someone to an approval queue in a system they open twice a week.
There's also a trust argument. Agents that act without prompting can feel opaque and alarming, especially when they're new. How much authority do you grant an AI agent before a human must sign off? The answer varies by context, and in customer service, the stakes of getting it wrong are immediately visible. Chat - because it requires a prompt - keeps a human in the loop at every step. That's not a bug for many regulated workflows. It's the point.
Technological revolutions are rarely binary. Transitions don't typically lead to total replacement. Instead, they create ecosystems marked by heterogeneity - a mix of old and new models, each finding its niche. Chat may be the wrong primary surface for agents, without being the wrong surface entirely.
What a better AI agent interface design looks like
Devin gets closest to an answer. The product panels the screen into browser, terminal, editor, and chat - four surfaces that make different aspects of agent behavior visible simultaneously. The panel layout acknowledges that conversation alone is insufficient. But the panels are presentation, not control surfaces. The user watches the agent work. That's still not enough: you can't queue approvals or inspect memory state through those panels. But it's directionally right.
What agents actually need from an interface is different from what chat provides. You are no longer designing a conversation. You are designing a policy: under what conditions does the agent act, what actions are permitted, what requires human approval, and how does the system communicate what it has done? That's closer to an ops runbook than a message thread.
The products winning in 2026 put the AI behind a verb, a canvas, a delegation, an ambient capture, or some sort of interactive output rather than behind a prompt. Ambient capture is the most underbuilt of these: an ambient capture watches what the user is already doing and produces useful artifacts in the background. The user does not invoke the AI at all. The AI is paying attention to work the user was going to do anyway. It produces transcripts, summaries, calendar events, or action items as a byproduct. That's not chat. It's closer to a capable observer who drafts the follow-up without being asked.
A teammate like Beagle approaches this problem from the other side: rather than waiting for a question, it can surface context, summaries, and handoff notes inside Slack and Teams as a byproduct of the work already in flight. The interface is wherever the work happens - not a separate prompt surface.
The Claude Routines trace timeline is the pattern worth watching. The review surface is a trace timeline: users can review what the agent did after the fact. OpenAI Codex runs headless in the cloud and returns diffs. Both are moving away from chat as the primary surface for agent output, even as chat remains the input. That split - chat for tasking, non-chat for reviewing - is probably where teams with mature agent workflows land.
The question isn't whether to use chat at all. It's whether to treat it as the whole interface. For agents that act over time, across systems, without being prompted: chat is a starting point, not a destination.