FINDING DIRECTIONAutomation can take over our tasks, but we are still responsible for the outcome.
I was driving through a construction zone where cones redirected traffic around the work area. My family was in the car when, all of a sudden, the steering jerked us toward the cones and a construction ditch only inches away. The lane-keeping assistance appeared to follow the painted lines instead of the temporary route marked by the cones. I disabled the feature almost immediately and have kept it off.
I think about that moment when I work with coding agents.
I use Codex and Claude Code. I prefer Codex for one behavior: I can redirect it while it is still working.
I can give Codex full access and let it work independently, but I still watch it. I follow the files it reads, the assumptions it makes, and the direction it takes. When it drifts off course, I send another message. The instruction enters the work already underway.
Codex calls this steering. A message can join the active run without cancelling it or starting another turn. Claude Code also accepts corrections while it works, but Anthropic describes its behavior differently: Claude stops what it is doing, then adjusts its approach using the new instruction.
That sounds like a minor interface difference. In practice, it changes how closely I can participate in the work. When I correct Codex, it does not discard the run I have been watching. I can intervene early, add context, and let it continue.
Fooling the Camera
Tesla says its Full Self-Driving (Supervised) feature requires the driver to remain attentive and ready to take control. An in-cabin camera watches the driver while the feature is active, and Tesla’s owner’s manual says the monitoring cannot be disabled.
Drivers have found another way around it. In June 2026, MotorTrend reported that some Tesla owners in China were mounting small plastic doll heads in front of the camera. The camera sees the doll instead of the driver, allowing supervised driving to continue while the person looks elsewhere.
The behavior predates camera monitoring. In 2018, NHTSA ordered the seller of Autopilot Buddy to stop distributing its product. The device prevented warnings that appeared when a driver’s hands were not detected on the steering wheel.
[ FIELD NOTE / ATTENTION ]
A camera is easily fooled. A prompt doesn’t guarantee the user understands the action.
The Approval Habit
The stakes of a software task and a moving car are radically different. Both interfaces use a simple response as evidence that a person is paying attention.
Agents still encounter moments when they need explicit permission. They ask before reading a protected file, connecting to the Internet, running a command outside a sandbox, or changing something that may be difficult to undo. Each prompt gives the person a chance to inspect the action and refuse it.
Permission dialogs interrupt us while we are trying to get something done.
[ FIELD NOTE / FRICTION ]
Clicking Allow is like accepting new terms and conditions no one reads. The prompt is another friction point we clear to keep moving.
Anthropic saw this happen in Claude Code. Its engineering team reported that users approved roughly 93% of permission prompts. The company found that people paid less attention as the prompts accumulated. It introduced automatic approval for safer actions partly to reduce that fatigue.
Watching the Work
The amount of work agents complete between decisions is growing. Anthropic analyzed millions of interactions and found that its longest Claude Code sessions nearly doubled in three months, from under 25 minutes to more than 45. Full automatic approval appeared in roughly 20% of sessions among new users and more than 40% among experienced users.
Experienced users interrupted Claude more often, shifting from constant approval to selective intervention.
That is closer to how I work with Codex. Full access removes repeated permission interruptions. I continue to supervise the task. I read the commentary, watch which tools it uses, inspect the direction it is taking, and step in when something does not look right.
Claude Code lets me intervene too. Its stop-and-adjust behavior makes that correction more disruptive in my workflow. Codex lets me add information to the active turn, which makes it easier to correct a weak assumption before it shapes the rest of the task. That single behavior is a meaningful reason I prefer it.
Access and supervision are separate choices. I can give my agent broad permission and remain involved in the work.
I stay close enough to see when the agent drifts and correct it while the work is still moving.
How much control do you give your agent, and are your hands still on the wheel?