How Atlas App Was Built: One Night’s Evidence Note

On the evening of October 19, 2026, Yusuf Aziz and three AI systems built a single-window workspace in about 16.5 hours where they could work together: Atlas App. The story itself is in the main essay. This page holds only that night’s numbers and how they were audited.

1. Why did it start?

Working on Antigravity’s web interface from a phone meant waiting for it to re-sync on every tap, with frequent errors. In Yusuf’s words: "That was the real trigger." The second problem was quota. Yusuf summed it up like this: "Frankly, what burns quota is not writing handovers, it is reading them." Having one agent read the long handover note left by another was simply too expensive.

2. Who did what?

Claude handled conversation and task coordination. Gemini (Atlas) wrote and ran the code. Grok served as an outside eye reading drafts from the shared board and pointing out flaws. Grok was given this role after it read a long industry article in one go on the evening of October 19 and found the same flaws Yusuf had noticed.

3. The handover test: How accurately does one agent read another’s note?

Four models were tested on 20 questions—half on an up-to-date handover note and half on an outdated note (`source: log measurement, 8 runs / 20 questions`):

• Gemini: 19 out of 20 correct, zero fabrications.
• Claude: 18 out of 20 correct, zero fabrications, but used roughly 2.5 times as many tokens as Gemini.
• Local model qwen (27B): 18 out of 20 correct. When given the outdated note, it refused to answer any of the questions on the outdated note.
• Local model gemma (31B): 16 out of 20 correct. It mistook three outdated records for current state.

Lesson from the test: Information passed between agents must stay small. That is why every handover now uses a short summary of at most 4 KB plus the last 10 lines.

4. What was built overnight?

• An identity-locked foundation so no agent can write a reply in another agent’s name.
• A single-column mobile screen unlocked with a single-use login code.
• A stream that does not lose messages when the connection drops.
• Dedicated native apps for macOS and Android.

By dawn the project had roughly 23,500 lines of code across 61 files (`source: git log count, 23,453 lines`). Training a small local model overnight was also tested; it summarized well, but because it fabricated a few details when reading handover notes, it was not put into live service.

5. The issue that surfaced in the morning

The live stream dropped five times. The models were not crashing: self-tests were touching the live server, and fixed timeout limits were shorter than real response times. Tests were moved to a separate port and the timeout floor was set to adjust automatically from recent response times. No drops occurred during a full hour of continuous monitoring afterward.

6. How was this page audited?

Every claim here was checked against primary logs by a separate auditor agent (Reaper) that did not write the text. Across 37 checks—16 technical measurements, 10 verbatim Yusuf quotes, 6 security checks, and 5 structural checks—no errors were found (October 20, 2026).