AetherOS is a web desktop for AI agents. The entire project sits inside a single HTML file. It starts with a boot screen. Then it opens into a dark workspace. The background has black animated particles. Everything runs on plain CSS and vanilla JavaScript. The site loads zero external dependencies and zero sound files.

The bottom dock holds seven apps. These tools manage four AI agents: Atlas, Sage, Bolt, and Vega. The Agents app shows them chatting and passing tasks. Files lets you browse a drive. Memory shows a live graph of what the system knows. Users can also modify a history timeline. Studio lets you test prompts. Flows builds node pipelines, and Cost tracks token usage.

Every window can be moved, resized, minimized, or maximized. The top menu bar has a model switcher. It routes tasks to Claude Opus 4.7, GPT-5.2, Gemini 3 Pro, DeepSeek V4, or Llama 4. The OS also has Spotlight search, sound toggles, and local API key inputs. The interface uses dark glass panels with blue highlights.

Building the node graphs in one single HTML5 file was tough. At first, the physics nodes pushed apart too hard. They floated right off the screen. Dragging extra lines while getting zoomed in also threw off the precise positions. I fixed the center gravity math. Then I divided mouse positions by the zoom factor. That stopped the unnecessary movement and kept the OS stable.😊

Behind the scenes on how our build process actually changed this month.

For most of this year we ran AI under a budget constraint, and that constraint quietly shaped everything. Which problems we pointed AI at. Which we did by hand. How much of a codebase we were willing to let a model read. Scarcity was the organising principle and we had stopped noticing it was there.

Then it lifted. Here is what actually broke.

The problem nobody warned us about. Freed from budget pressure, sessions got long and deep. A model will happily work a huge problem in one sitting. But the sitting ends — timeout, full context, or you go to bed. And everything it worked out along the way vanishes. The dead ends it ruled out. The constraint it discovered in hour two. The reason it chose one approach over another.

When sessions were short, we held that context in our heads. When they got long, we couldn't. We started losing more to forgotten context than we had ever lost to rate limits.

What we built to fix it:

Exit interviews. Before closing any meaningful session, we prompt the model for a handover briefing — what it did, what it decided, what the next instance needs. Not a commit message. An actual briefing. That output becomes the opening context next time. This is the single highest-leverage thing we changed and it took an afternoon to adopt.

Browser session stability. Long builds kept dying when whoever was driving switched tabs. Moving active sessions into the Chrome extension fixed it. Unglamorous. Recovered more hours per week than any clever prompting.

Logic only, no UI. We kept fighting the same battle — strong architecture and logic coming back attached to cluttered, incoherent interfaces. So we stopped asking. We now prompt for logic and architecture and explicitly block frontend generation. Structural foundations go to a human who builds the interface natively. Restricting the model made the output better, which is not what I expected.

Two-stage handoff. Fast rough first pass to kill ambiguity, often straight from a client scope document — one was 125 pages fed through in a single stretch. Then a different person takes the careful second pass.

The part I did not expect. More capacity meant more verification, not less. We increased the QA burden at the same time we removed the ceiling, deliberately, because an agentic system will always find something to report and the report will always sound like progress.

The tools were the easy part. The reorganisation was the work.

Full write-up: https://infiniteup.dev/the-bottleneck-moved-how-we-rebuilt-our-workflow-around-mcp-and-fable/

budget shaping the whole process, good honest writeup