Agent Controls Need Weight

There is a particular kind of office object that makes a person feel briefly more competent than the day has earned. A nice keyboard. A well-weighted dial. A programmable button with a label you chose during a burst of optimism and now press like you are operating a small, private machine room. It is funny how quickly a tool becomes part of the body once it has a satisfying click.

So it was only a matter of time before AI agents acquired hardware. Business Insider reported that OpenAI has teased a Work Louder-built device for Codex, apparently called Codex Micro, with a July 15 reveal and the promise that Codex shortcuts are getting an upgrade. The details are thin, and the sensible adult response is to wait for the actual product. Still, the image is useful: a little grid of controls for a system that can read a repository, edit files, run commands, and return work for review.

Codex was introduced as a cloud-based software engineering agent that can work on many tasks in parallel, each in its own sandbox, returning terminal logs and test outputs so a person can inspect what happened. GitHub's Copilot cloud agent follows a similar pattern: it can work autonomously in a GitHub Actions-powered environment, research a repository, plan, change code, and open a pull request. A recent paper, The Shift to Agentic AI, found that more than 10% of Codex users managed three or more concurrent agents in a week, while reusable skills and longer delegated tasks became more common.

That is a lot of background work to supervise with a chat box and a notification tray. The obvious interface answer is shortcuts. Start this. Retry that. Explain the diff. Open a PR. Run tests. Cancel. Escalate. Approve. If the agent is becoming a workbench, the workbench gets controls. People who do repetitive, high-context work have always built little rituals and muscle-memory paths around it. The developer has a command history full of incantations that would look concerning to a normal person but are really just Monday.

Control surfaces matter because work is physical even when the material is digital. A button can make a decision feel real. A dial can slow the hand down just enough for the mind to arrive. A glowing key can say, with more honesty than a tiny badge in a sidebar, that something is still running.

The risk is that agent controls become speed controls only.

If every new button means "do more agent work with less friction," the interface quietly teaches the operator that supervision is another throughput problem. Launch three sessions. Approve the small patch. Ask one agent to draft the migration plan while another hunts the test failure. It is intoxicating in the specific way competent tools are intoxicating: the clean feeling that the work is finally moving at the speed of your impatience.

Anyone who has managed several agent sessions knows the other feeling. The screen looks productive and the room feels faintly haunted. One agent is done but needs review. One has changed a file you forgot was coupled to a customer-facing flow. One is still running tests. One produced a plausible plan that may be wrong in the expensive middle. Nothing is on fire. Everything is asking for a little judgment. The person has not been replaced by the agent. The person has been turned into a small air-traffic controller for unfinished intentions.

The shape of control matters here. A good agent button should make consequence visible. What system will this touch? Which credentials, files, tickets, customers, or budget meters sit inside the blast radius? Is this reversible, reviewable, live, or an expensive question wearing a cheerful icon? What evidence will come back, and what will still require taste rather than a green check?

The weight of a control is not just the resistance of the switch. It is the social and cognitive friction around the act.

We already know how bad weightless controls can get. The consumer web spent two decades making decisions feel like taps, then acted surprised when people subscribed, consented, purchased, shared, and regretted things with the same thumb motion. Enterprise software did its own version with approvals: a box gets clicked, a workflow advances, and somewhere downstream a decision has acquired the moral authority of process.

Agents raise the stakes because the button can trigger work, not merely record preference. The agent can inspect, alter, summarize, commit, buy, schedule, message, route, and retry. It can keep working after the person has moved on to another window, another meeting, another mood. A control that feels like a shortcut may actually be a delegation contract.

That contract deserves a better interface than vibes and velocity. Good agent controls should probably be less like media keys and more like cockpit instruments, though without the tragic amount of beige plastic. They should show active work as a first-class state. They should make stop and pause feel as available as start. They should distinguish a harmless local draft from a task that will spend credits, consume CI minutes, alter shared files, or create review debt for someone else. They should preserve the trail without turning every user into a compliance clerk before lunch.

They should also respect the fact that human attention is not infinitely parallel just because agents are. A person can launch five tasks, but she cannot form five careful judgments at the same instant. The interface should help pace the return of work: grouping related results, surfacing conflicts, marking stale context, and refusing to flatter a crowded queue as evidence of mastery. Sometimes the humane control is not "run all." Sometimes it is "make me look at this before I create more of it."

The deeper shift is that AI interface design is moving from asking to steering. The prompt box was a conversation starter. The agent control surface is a responsibility surface. It decides which actions feel normal, which pauses feel natural, and which kinds of oversight get reduced to a satisfying click.

I am not against the click. I like a good button. I have rearranged a desk around less. But the next generation of AI controls should earn their tactility. If a tool can act in the world of work, its controls should help people feel the weight of that action before the machine runs off to be helpful. Agents will be shaped by model capability, sandbox policy, and benchmark curves, yes. They will also be shaped by the small human moment before the button is pressed, when control either becomes responsibility or dissolves into convenience.