---
title: "Project Atlas explained"
slug: video-project-atlas-explainer
section: Video Library
status: stable
description: "Every time we ship code, someone has to ask the same question."
last-updated: 2026-10-07
---

# Project Atlas explained

Video: Project Atlas explained (2:00) — /docs/video-project-atlas-explainer

- Answers: What is Project Atlas and what are its limits
- Audience: engineering; program scope: all programs
- Visibility: staff; last verified 2026-11-02; owner platform-team

Chapters:
- 0:00 The problem: did we break something — /docs/video-project-atlas-explainer?t=0
- 0:06 What Atlas is — /docs/video-project-atlas-explainer?t=6
- 0:14 The principle: agents recommend, humans decide — /docs/video-project-atlas-explainer?t=14
- 0:26 How Atlas thinks — /docs/video-project-atlas-explainer?t=26
- 0:36 How a run works: six stages — /docs/video-project-atlas-explainer?t=36
- 0:55 When it runs — /docs/video-project-atlas-explainer?t=55
- 1:01 The hard limits — /docs/video-project-atlas-explainer?t=61
- 1:19 Where it stands — /docs/video-project-atlas-explainer?t=79
- 1:35 What is next — /docs/video-project-atlas-explainer?t=95

## Transcript

Every time we ship code, someone has to ask the same question. Did we just break something? Meet Atlas, Alliance Abroad's advisory testing agent built and owned by Platform Engineering. Its principle is simple. Agents recommend, deterministic checks prove, humans decide. Atlas reads the Nexus documentation and uses the Claude API to design test scenarios. But the model only ever writes text. It never executes anything. Every scenario lives in Kiwi TCMS and only cases a named person has confirmed are allowed to run. Every run follows six stages, fetch the confirmed cases, lint them against a strict contract, run them in a real browser as a dedicated bot account, judge the outcome with deterministic assertions, not the model, triage any failure, and report to Platform Pulse and Kiwi. Atlas runs every weekday morning, after each deploy to a development slot or on demand. And it has hard limits. Development only, never production. It can't block, approve, merge, or promote anything. It never uses a real person's account, never touches payments or service records, and never sends personal data to the model. So how is it doing? Three test plans covering 49 cases, two real product defects found, and just over $5 of model spend in October. The first formal go-no-go review said no-go as frozen, so Atlas stays advisory, earning trust one run at a time. What's next? A release confidence pilot, proving one real outcome end-to-end. A staff edit to an applicant's contact details is saved, audited, and changes nothing else. A decision on Atlas' future follows the 30-day review, closing November 6th. Project Atlas. Agents recommend. Deterministic checks prove. Humans decide.