Skip to content
Case study "chat documents" web app (RAG)

The atelier where AI agents
craft your software

This is how one idea became a live, tested application, with AI agents doing the work and people making every call.

Watch the 4-minute story
0:00 / 4:41
The story · case study

One idea. Nine AI agents. A live app.

Watch a plain-language brief become tested, live code.

4:41CaptionsSound on
It started with a brief

One plain-language paragraph

New project brief
Build a chat application where users can upload documents and ask questions about them.
The flow

One brief in. A live app out.

Nine agents in sequence. A person approves at every gate. Wren documents every stage. Every integration goes through MCP.

  1. Brief: input
  2. Alex: requirements
  3. A person approves
  4. Archie: architecture
  5. A person approves
  6. Piper: planning
  7. Maya: app design
  8. A person approves
  9. Devon: development
  10. A person approves
  11. Tessi: testing
  12. Otto: infra
  13. A person approves
  14. Dex: deployment
  15. Live app: output
  16. Wren documents every stage
  17. Jira, GitHub, AWS and Gemini connect through MCP
Fig. 1 · The Atelier lifecycle
STAGE 0/9APPROVED 0/5DOCS v0MCP SYNC 0/4
agent at workapproved by a personfailure, then fixed
The story

Nine agents. People make every call.

Each agent does one job, then hands over — but only after a person approves.

01 · Requirements

Alex, requirements analyst

Instead of guessing, Alex interviews the team: five focused questions, each with a recommended answer and the reasoning behind it.

People stay in control. Alex recommended a shared password; the team chose an open link, so Alex adapted.

The team picked the lightweight scope and tuned the stack: React 19, MinIO, no Redis.

0findings
0epics
0user stories
0acceptance criteria
Approved by a person
Question 3 of 5Alex · interview
How should people get into the shared workspace?
Shared passwordRecommended
Open linkTeam's choice
Alex adapted: added per-IP rate limiting and a document cap to protect the open link.
02 · Architecture

Archie, solution architect

Archie designs the system from the approved requirements: six components, from the web app to the vector database and Gemini.

A sequence diagram for every critical path, with confidence and cost on every output.

0components
0API endpoints
0data entities
Approved by a person
System designconfidence + cost on every output
Web app
API
Ingestion
MinIO storage
Vector DB
Gemini
10 endpoints5 entitiessequence diagrams
03 · Planning

Piper, planner

Piper turns the design into six sprints, each item estimated and assigned to the right specialist agent.

Every item is pushed into Jira automatically and kept in sync both ways.

0work items
0sprints
0sync failures
Approved by a person
Sprint plan93 items
S1
S2
S3
S4
S5
S6
Jira synced both ways · 0 failures
04 · App design

Maya, designer

Maya proposes three design directions. The team picks Shared Noticeboard, and Maya builds the brand around it.

Every screen is a live, working preview, and the code repository is scaffolded from it automatically.

0directions
0chosen by the team
Chosen by the team
Design directionslive previews
Direction A
Shared Noticeboard
Direction C
✓ picked by the teamrepo scaffolded
05 · Development

Devon, developer

Devon codes all six sprints, committing to GitHub and updating Jira, with a person approving every sprint.

When checks fail, it runs a fix pass until they pass. A final review across all sprints catches the bugs in between.

0tickets delivered
0commits
~$0AI cost
Approved by a person
Sprint 4 · feature/upload-limitsGitHub + Jira
› commit 7f3a1c2 add per-IP rate limiter
› commit 91be04d enforce document cap
✗ checks failed: 2 tests
› fix pass 1 correct limiter window
✓ checks passed
✓ sprint 4 approved by a person
06 · Testing

Tessi, tester

Tessi writes 325 unit tests, runs the end-to-end suite, and catches a real bug before release.

0unit tests
0real bug caught
Approved by a person
Test runend-to-end ✓
0unit tests passing
E2Eend-to-end suite green
Real bug caught before releasefixed and re-tested
07 · Infrastructure

Otto, DevOps engineer

Otto plans a minimal AWS setup from the approved design, with an estimated monthly cost before anything is created.

After launch, Otto reviews spend and suggests where to downsize.

Approved by a person
AWS resource planestimated before creation
Container service web + API
Object storage documents
Vector database embeddings
Load balancer public link
Savings ideas after launch
08 · Deployment

Dex, release engineer

Dex ships it to AWS. Every failed attempt came with a clear diagnosis, and the fixed version went live.

Approved by a person
Release · AWSlive
› deploy attempt 1
✗ failed: diagnosis attached
› fix applied from diagnosis
› deploy attempt 2
✓ health checks passing
✓ live on AWS
09 · Documentation

Wren, technical writer

Wren updates the documentation after every stage, down to a trace from each requirement to the live code.

0document versions
Updated after every stage
Documentationupdated every stage
v1v2v3v4v5v6v7v8
Requirement FR-012StoryCommitLive code
The result

And here it is, live

The app the team described, answering from their own documents, with sources.

Designed by MayaBuilt by DevonTested by TessiShipped by Dex

Illustration of the delivered app. The real app is shown at the end of the story video.

Metrics

Every step measured

Cost, output, quality and governance, tracked for every project. This is what the case study added up to.

AI cost

From one brief to tested, live code
~$0
for the whole lifecycle, every stage included
Development ~$35All other stages ~$15

From brief to backlog

What Alex and Piper produced

Build

Devon, sprint by sprint
0tickets
0commits
0sprints
Sprints approved by a person6 / 6

Quality

Tessi's test run
All passing

End-to-end green

1 real bug caught

Governance

People in control, everything recorded
Human approval gates5 / 5
Documentation versions8
Jira sync failures0
Synced both ways
sample record

Every AI output carries its receipt

Shown on every requirement, design, plan and commit in Atelier
Model
Gemini
Prompt version
req-analyst v3.2
Execution time
4.8 s
Tokens
12,480
Cost
$0.06
Approved by
a person ✓
Every number here is tracked live in Atelier, in Analytics and Governance.
Integrations

Works with the tools you already run

Jira, GitHub, AWS, Azure and more — kept in sync, both ways.

Every connection goes through MCP. Agents never touch your systems directly.
From one brief to tested, live code, for around fifty dollars of AI, with people approving every key step.