ProjectBuild an Agentic Content Studio: the explanation · Module 2: The solution: one prompt, three agents, one spec · part 4 of 22
Part 4 · Module 2: The solution: one prompt, three agents, one spec

Module 2: The solution: one prompt, three agents, one spec

21 min read·
About 45 minutes · Needs: Module 1 (the problem) · Code: backend/app/ai/pipeline.py, backend/app/ai/jobs.py, frontend/src/render/renderer.ts

Where this fits

WHERE THIS FITS Slidesmith Studio this module BROWSER React FRONT DOOR AND API nginx, FastAPI BACKGROUND AGENTS LangGraph MODELS Ollama RENDERING Playwright DATA Postgres 17 OPERATIONS evals, Docker, Cloud Run Studio app React pages /render page same renderer nginx proxy, headers Auth sessions, CSRF Brand kit onboarding, uploads Projects deck spec, edits Job runner progress over SSE Research search, facts Strategist outline Designer slide JSON Check fit, else repair LangGraph pipeline state graph, JSON repair Ideas post ideas from pillars qwen3.8:27b Ollama, local GPU Web search Ollama API Chromium render, measure Exports PDF, PNG, ZIP, HTML PostgreSQL users, decks, jobs Uploads disk or GCS Evaluation suites, gate Insights usage, timings Containers Docker Compose Deploy Cloud Run, CI/CD
The whole generation path is highlighted: the studio starts a job, three agents write the deck with the model (amber), Chromium checks it and the exports leave the app.

Module 1 (the problem) showed why a model should not write the finished slides. This module is the answer in one pass: what happens between the moment Northwind Labs types a prompt and the moment a PDF downloads. Every later module zooms into one box of this picture, so it is worth getting the whole path clear first.

The rest of this project is yours to keep

This project is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Or open everything with a plan