Derek Cheng

The Two Zoom Levels of Agents

Building @Tonkotsu has taught us many lessons in product design, with both the underlying technology and user behavior shifting rapidly.

One of the hardest but most critical lessons we’ve learned is about calibrating the zoom level — how close or far the user feels from the work. You can see examples of this play out across the industry:

  • Codex gets flak for going “heads-down” for too long compared to Claude. Users feel too zoomed out from the work.

  • By contrast, Cursor and IDEs are starting to feel too zoomed in. When the majority of code is written by agents, an editor-first UI is a misfit.

We've had zoom level mismatches with Tonkotsu as well, and when we dug into it, the usage data was contradictory, suggesting that users wanted us to both zoom in and zoom out. We scratched our heads a bit at this seeming contradiction, but it clicked when viewed through a management frame: managers delegate at a high level but want updates at a fine-grained level.

The key lesson: there are two different zoom levels that agent builders need to intentionally design for. Actions need to be coarse enough that the user can be efficient and productive. But observations need to be fine-grained enough that the user is confident that progress is being made.

More here → https://blog.tonkotsu.ai/p/the-two-zoom-levels-of-agents

I'm curious what you are seeing play out in your experience, as either users or builders of agents? What products feel too close vs too far vs just right?

51 views

Add a comment

Replies

Best
AJ

The dichotomy is false most of the time. I don't want either, each stage of the process gets different zoom levels.

If I'm running a security audit I want details, I want line numbers, snippets, diagrams, I want to deeply understand why something is vulnerable. It's context dependent and situational all the way down.

For work that is well specced, I don't need that granularity in coding, I need the tests to be faithful and passing, I need the end to end flow to work.

What we want is the ability to slow down and have detail on demand for when we need it. It's not binary either. it's a spectrum.

We are constantly traversing from the general to the specific and back up again. In this comment alone you can see it. First sentence is a general statement, second and third have specific examples, then I generalize again, and now this meta comment goes up a level of abstraction.

Claude code is very customizable, you can choose at what level you're at and how much detail you want. but you manage that manually.

Opencode is similar.

Tonkotsu takes an opinionated approach. Giving you structured zoom levels, discrete steps instead of a slider. I think a lot of people find value in this. The structure is the point, Tonkotsu lends itself to self documenting specification, your spec doc is your implementation and the feedback is specific where it needs to be. It fits a certain profile of person very well.

Derek Cheng

@build_with_aj Appreciate your thoughtful response! It's a good point that not only should zoom level vary by stage of the process (delegation vs observation of status), but also by task type! So much to build here...

AJ

@derekattonkotsu 

I'm gearing up for a huge project and I am tempted to see just how far Tonkotsu can go. It's an ambitious thing and I've been using opencode. Mostly because my workflow is more iterative.

But I am tempted to see if Tonkotsu can handle some downright lunacy in terms of scope.

Actually. I recently wrote on how I do qualitative benchmarks for models, so I can select models for coding.

I want to test one of the benchmarks to Tonkotsu and I will share the results. BTW, I hope you ahve an opensuse package because I ditched windows!

Derek Cheng

@build_with_aj Yes, Tonkotsu's model is definitely more hands-off delegation. Would be really interested to see how it does on that project!

In terms of Linux support: on our radar, but presently we only support macOS and Windows.

AJ

@derekattonkotsu Aight,

Wine and windows it is! I'll probably test this weekend. See if I can get it running on WINE or bottles.