I was in a Claude Code session this week. The agent was deep in a multi-step workflow, working through a plan. A new thought hit me, the kind that's useful three steps from now but worth saying before I forget. I typed it in mid-stream, the way you'd interrupt a colleague who's already half a sentence into something. Claude wrote back: "Got it, I'll research that when I get to the seed step so the scenarios are authentic to this situation. Continuing with my current work, and we'll get to that next."
The Feel Changed
That kind of mid-task acknowledgment isn't new. Claude Code has supported sending updates while the model is working for a long time. So has the deep research feature on ChatGPT. You could always do this. What's new is that it has started to feel right.
I don't know exactly what shifted. The response itself feels different. The acknowledgment used to read like a confirmation receipt, and now it reads like a colleague telling me they heard me, where they'll fold the note in, and why. Reliability is up too. The next step actually takes the mid-flight input seriously in a way that wasn't always true. Some of the shift might just be me, getting comfortable with a pattern I used to second-guess. Whatever the mix is, the whole loop runs cleaner than it did six months ago.
This changes how you work alongside the model. The default pattern with an AI that runs long jobs has always been to wait. You hand it the task, you sit on your hands, and when you remember a piece of context twenty seconds in, you have a choice. You can stop the run, edit the prompt, and start over. Or you can stash the note in another file and copy-paste it back in when the current job finishes. Both options break your train of thought, and both punish you for thinking of something useful at the wrong time.
The version where you just type the thought into the session as it occurs to you is a much smaller cognitive ask. You don't have to track where you are in the run. You don't have to context-switch into a notes app. The session itself becomes the place where the work and the side thoughts live, and the model handles the splicing.
Two Companies Moving the Same Way
ChatGPT has been working on the same thing inside Deep Research. The early version of that feature was opaque. The model started its long-running task and you watched a loading bar. Now you can drop instructions in while it's still going, and the next phase reflects what you said. The mechanics differ from Claude Code's, but the design intent points the same direction.
I'm using both. I notice the same trend in both. Two companies, with different products and different goals, getting the same problem right at roughly the same time. That's worth paying attention to, because it suggests the pattern is settling. The expectation that you can update the AI mid-run and have it reasonably incorporate what you said is becoming table stakes for any agent that runs longer than a single turn.
Cowork, on the Claude side, isn't fully there yet. When you drop a thought in mid-session, the behavior is inconsistent. Sometimes the agent recognizes it and picks it up after the current step finishes. Sometimes it goes unacknowledged. Sometimes the message ends up stuck in the queued-message UI and never gets addressed at all. I'm sure they'll fix that soon if they haven't already. For now, knowing what Claude Code can do makes the gap easy to notice.
This kind of improvement doesn't show up in a launch announcement. There's no headline, no version bump, no demo video. It's a feel thing that arrived in pieces over months, and the only way you'd catch it is by using the tools day after day. For anyone leading a team on these tools, that's the actual texture of how the products are improving right now. The reshape of knowledge work is going to keep happening like this. Quiet quality-of-life shifts that change the rhythm of how you work, layered into products you're already using. The people who feel it are the people putting hands on the tools.