Transcript
00:00:00For two years, we've been handing our code bases to AI agents and letting them rewrite
00:00:05whatever they please, and then reviewing that by just scrolling a wall of code.
00:00:09Somebody finally got fed up with that and built the opposite, a way to actually watch
00:00:14an agent move through your code in 3D.
00:00:17It's called MindWalk.
00:00:18Now, this might even be a better way to visualize our code and bring it to life.
00:00:23We have tools like Graphify, so how does this even compare?
00:00:30Now, here's the real problem MindWalk is aimed at.
00:00:34Our agent's log tells us what it did, ran this, edited that.
00:00:38What it never tells us is how it understood the job.
00:00:41Which files did it decide were relevant?
00:00:44Where did it go looking around before it changed a single line?
00:00:47Did it stay in its lane or quietly just go through half the repo?
00:00:51All of that.
00:00:52The agent's actual thinking is kind of invisible in a transcript.
00:00:55It's just gone.
00:00:56If you enjoy coding tools to speed up your workflow, be sure to subscribe.
00:00:59We have videos coming out all the time.
00:01:01So let me show you what it looks like when it's not gone.
00:01:04I've got a project here in a Cloud Code session I ran earlier.
00:01:08I can go install this, go, no pun intended, and then build it out.
00:01:14Now I can fire up MindWalk with this command here, and my whole code base renders as this
00:01:20dark 3D city.
00:01:21Every file is a building, and the more code it has, the taller it stands.
00:01:26Then I press play on the session.
00:01:29Watch the lights here.
00:01:31A file glows green, mossish green, when the agent just glances at it.
00:01:36Blue when it actually sits down and reads it, and amber when it makes an edit.
00:01:41And I can drag the timeline back and forth and literally trace the path the agent walked through
00:01:46this code.
00:01:48It's one go binary that just reads the session logs, clawed code, and codecs are already writing
00:01:53to your disk.
00:01:54Plus a 3JS scene in your browser for the 3D.
00:01:58That color language is the real trick here.
00:02:01Cool colors mean the agent was looking.
00:02:03Warm colors mean it was changing.
00:02:06The timeline's marked up too.
00:02:07Little symbols show you exactly where a sub-agent spun up, where the context window got compacted,
00:02:14and where you barged in with a new instruction.
00:02:17So now you always know where you are going.
00:02:20But here's the piece I actually think matters.
00:02:22Once you're passed, this looks cool, right?
00:02:24There's an analyze command that runs a judge over the whole session using your own local
00:02:29cloud or codecs, and it grades the run it just went through.
00:02:33Did the agent explore enough before it started editing?
00:02:37Okay, did it stay in scope?
00:02:38Did it wander?
00:02:40Did it bother to verify its own work?
00:02:42And every one of those findings is a clickable link straight to that moment on the timeline.
00:02:47That's a code review of the agent.
00:02:49And if you've got no session replay at all, there's a map command that'll render any repos
00:02:54as a 3D city on its own without any setup, which is just a really cool trick.
00:02:59Okay, now, here we go.
00:03:00When this thing launched on Hacker News, the number one comment was pretty honest,
00:03:06because you're probably asking it too.
00:03:08This is very cool, but would I actually use this?
00:03:10What would I actually use this for?
00:03:12And some people didn't even fight back on that.
00:03:14They said things like, it serves no real purpose.
00:03:16It's fun, and not everything needs a reason.
00:03:19Someone else eventually landed a fair jab for a small change.
00:03:23A plain old git diff, it's just faster than watching an agent go through this 3D tour.
00:03:28I get it.
00:03:29So, this is way too fresh.
00:03:31It's version 0.3.
00:03:32It only reads Claude code in Codex logs.
00:03:34If you live in Cursor, you can't replay your sessions yet.
00:03:38Which brings me to the question, who's this genuinely for?
00:03:41Is it just fun?
00:03:42Well, if you're comparing models, you want to see how Claude explores.
00:03:45A task versus how Codex does.
00:03:48This is a really cool lens you can't really get anywhere else.
00:03:51If you're staring down a huge sprawling agent change, and a flat diff just isn't giving you
00:03:56the shape of what happened, this 3D fly-through plus the judge report is pretty cool.
00:04:01It breaks that down.
00:04:02But for a two-line fix, come on.
00:04:04Just read a diff.
00:04:05I don't think MindWalk is going to replace any of our review tools.
00:04:09But it's here, maybe, to give you a feel for the run that your tools were never built to
00:04:14actually show us.
00:04:15And that's the part that we can actually play around with here.
00:04:19We taught our machines to write our code way faster than we ever built a way to understand
00:04:24what they wrote.
00:04:25MindWalk, like graphify, graphify, graphify.
00:04:29Okay, I think I'm saying that right.
00:04:31This could be a first serious swing at the gap, letting you see how an agent actually reasons
00:04:36across the code base, not just the keystrokes it left building it out.
00:04:40It's rough, it's early, and it's aimed at the exact right problem.
00:04:43But play around with it, see if it even helps you.
00:04:46If you enjoy coding tips and tricks like this, be sure to subscribe to the BetterStack channel.
00:04:50We'll see you in another video.