Before
The loop most of us write
It works in the demo. In production it retries instantly, forever, on errors that will never succeed.
Short, interactive lessons called bytes that teach you to build AI features that hold up in production: retries, caching, structured output, agents and guardrails.
Byte 01 is free to read now. The interactive Track 1 opens soon. Get Byte 01 free
Scenario: an AI agent issues a refund through a payments API that supports idempotency keys.
What's a byte?
Every byte has the same shape, so you always know what you're getting and where you are.
Byte 01, line by line
Before
It works in the demo. In production it retries instantly, forever, on errors that will never succeed.
Step 1 of 5
429s, 5xx errors and timeouts can succeed on a second try. A 400 or 401 never will, so fail fast and say why.
Step 2 of 5
Four tries, then surface the error. An outage shouldn't hang your app or your user.
Step 3 of 5
Wait 1s, then 2s, then 4s, and randomise each wait so a thousand clients don't all retry in the same instant.
Step 4 of 5
When the API tells you when to come back, listen. It knows its own load better than you do.
Step 5 of 5
Your SDK already retries. Turn that off, or drop your loop. Stacked layers quietly turn one request into twelve.
Scroll, tap a step, or use the buttons.
Byte 02 preview
Most apps resend the same system prompt and documents on every call. Move the sliders to match your app.
You'd spend less on input tokens
–
saved every month
Simplified: input tokens only, assumes every request hits the cache, ignores cache-write costs and output tokens. Byte 02 covers how to structure prompts so the cache actually hits.
Start where you are
Three quick questions, each about one topic. Track 1 below then highlights where to focus. It's a pointer, not a test.
Track 1 · 6 bytes · about 42 minutes
Six bytes, from basic to advanced. Each one ends with something you build and get checked.
Retries that don't make outages worse Free
You'll build a retry wrapper that survives 429s and 529s.
Field noteRetry storms that turn a small outage into a big one.Prompt caching to cut cost and latency
You'll restructure a prompt so repeat calls cost less and return faster.
Field noteThe same 4,000-token system prompt, paid for on every call.Validate LLM output and recover from failures
You'll add schema validation and automatic repair to structured output.
Field noteOne malformed response that stops a whole batch job.Why your agent loops forever
You'll give an agent step budgets, loop detection and a safe exit.
Field noteAn agent calling the same tool until the budget runs out.Idempotency: retry without double-charging
You'll prevent duplicate actions when a side-effecting call is retried.
Field noteFrom payments, where a duplicate retry is a duplicate charge.Guardrails every AI feature needs
You'll wrap one feature in input checks, timeouts and fallbacks.
Field noteThe input nobody tested, in front of a real customer.From production, not from tutorials
The patterns come from systems where failure costs money. Then they're cut down to seven minutes, so you get the lesson without the incident.
Each byte is built around a failure mode that shows up in real systems, explained by someone who has debugged them.
18+ years building payment systems: reconciliation, settlement, ledgers and idempotency, where a retry that runs twice charges someone twice.
Senior engineers sharing the patterns they rely on day to day, starting with Track 2.
Coming in Track 2Where this goes
Each track builds on the last, so the bytes compound.
Retries, caching, structured output, agent loops, idempotency and guardrails.
Multi-agent pipelines, skills, tool design and evaluations: agents that do real work and know when to stop.
Design reviews, failure modes and the trade-offs nobody writes down.
About seven minutes: a two-minute visual explainer, four minutes building the pattern yourself, and one minute of automatic checks.
Partly. Each byte opens with a short explainer, but most of the time you're writing code that gets checked, not watching.
Developers who can already code and want to build AI features properly, from early-career to senior.
Python examples first, with more to follow.
Byte 01 is free, no card needed. The full Track 1 is a one-off A$29, or A$19 for our first 100 founding members. No subscription: pay once, keep it.
Byte 01 is available now when you sign up. You'll get one email when the rest of Track 1 opens.
Get the Byte 01 pattern and code now, free. The interactive version and the other five bytes open soon: one payment of A$29 for all of Track 1, or A$19 for our first 100 founding members. No subscription.