Skip to content
NEW v0.2.0 — Tracks, roadmaps and tiered stacks
STACK IT FIRST
Install the skills
Browse the site
03 · Backend · Roadmap

Four questions, one stack.

A 4-question roadmap for choosing a backend architecture an AI agent builds well in. Every answer is explained; every path ends on a stack.

Interactive roadmap

Stage 1 of 2Where are you?

Does something in production already depend on this backend, with more than two engineers touching it?

A backend with no consumers can be rewritten by an agent in an afternoon. A backend with consumers needs a versioned contract, migrations that roll forward safely, and a test suite that fails before a client does.
Stage 1 of 2

Where are you?

Backend tiers are about how many things break when a contract changes. The more consumers an API has, the more the agent needs a schema it can diff and tests it can run before touching a handler.
Stage 1 · Q1

Does something in production already depend on this backend, with more than two engineers touching it?

Why this matters

A backend with no consumers can be rewritten by an agent in an afternoon. A backend with consumers needs a versioned contract, migrations that roll forward safely, and a test suite that fails before a client does.

yes
Stage 1 · Q2

Do you run multiple services, teams or regions, or sit behind a compliance boundary?

Why this matters

Each of these turns implicit coupling into an outage. Service boundaries, audit logs and data residency all need to be explicit in code, which is exactly what makes them tractable for an agent. Below that line, one well-structured service on managed Postgres is still the right answer.

yes Tier

Scale

Compiled, typed services behind explicit contracts, with observability built in.
no Tier

Growth

One service with modules, managed Postgres, a queue, and an OpenAPI contract.
no Tier

MVP

One process, one file-based or hosted database, a schema-validated router.
Stage 2 of 2

What are you building?

The profile decides what the backend is optimised for: request latency for a product, throughput and retries for data work, or speed of change for internal tools.
Stage 2 · Q1

Does this backend mainly serve requests from your own web or mobile app?

Why this matters

A product API is judged on p99 latency and on how little glue sits between the schema and the client. Sharing types end to end lets the agent change a field once. Anything else is optimised for a different axis, so the framework choice changes.

yes Profile

Product API

The API behind your own product, optimised for latency and end-to-end types.
no
Stage 2 · Q2

Is the core job moving or transforming data, such as ETL, embeddings, LLM pipelines or scheduled jobs?

Why this matters

Data and AI workloads want idempotent steps, retries and a language with the best library ecosystem for the job, which in 2026 is still Python. Internal tools want the opposite: a batteries-included framework where an agent can add a screen and a table in one change.

yes Profile

Data & AI

Pipelines, jobs, embeddings and LLM workflows, optimised for retries and throughput.
no Profile

Internal

Admin panels, back-office tools and automation, optimised for speed of change.
Result

Tier × profile.

Stage 1 picks the row, stage 2 picks the column. Each cell is a stack with a rule you can install.
Tier Product API serves web & mobile Data & AI pipelines & jobs Internal tools & automation
01 MVP Ship it Bun + Hono + SQLite 5/5 FastAPI + Pydantic + Postgres 5/5 Hono + SQLite + server-rendered HTML 5/5
02 Growth Charge for it NestJS + Postgres + Redis 5/5 LangGraph + FastAPI + Postgres 4/5 Django + Postgres + Redis 5/5
03 Scale Run it for many Go services + Postgres, Rust where latency is the product 4/5 ClickHouse + a typed ingestion layer + Postgres 3/5 Rails + Postgres + Redis 5/5