Integration Engineering
From your first API call to securing AI agents: forty short posts on connecting software that was never built to talk. APIs, every kind of auth, webhooks, retries, syncing data, enterprise needs and agents. In order from simple to advanced, and each one stands on its own.
Every post stands on its own. Start with any of them.
What an Integration Engineer Does All Day
A company uses many tools, and none of them talk to each other. Making them talk, safely, and keeping them talking is the whole job.
Read article →What an API Is: A Door With Rules Written On It
An API is a door one program opens for other programs. Every call to it answers the same four questions.
Read article →Reading an HTTP Request: The Four Parts Every API Call Has
Most API bugs are not hard. A request has the same four parts every time, and reading them in order finds most problems.
Read article →Status Codes That Matter: The Number Tells You Whose Fault It Is
Before you read any error message, read the number. The first digit tells you where to look, and whether trying again can help.
Read article →API Key or OAuth? It Depends on Whose Data It Is
For your own data, an API key is fine. For your users’ data, use OAuth: they approve it, you get a token, and you never see their password.
Read article →Five Ways an API Asks “Who Are You?”
None, API key, Basic, Bearer and OAuth look alike in the docs, but each fits a different job. Here is how to tell which one you are holding.
Read article →OAuth in Five Steps, Without the Big Words
Behind most “Allow access” buttons is the OAuth authorization code flow. Here are the five steps, and the two small things that keep it safe.
Read article →My LinkedIn Token Dies Every 60 Days, On Purpose
Access tokens expire so a leaked one stops working on its own. Here is how I handle one that lasts 60 days with no refresh token.
Read article →When No Human Is There to Click “Allow”
A nightly job needs a token at 2 AM and nobody is awake to log in. The OAuth client credentials flow lets your server get a token for itself.
Read article →No API Gives You All the Data in One Call
Data comes in pages, and your code has to ask for the next one. If it does not, it quietly loses records in real use.
Read article →429 Does Not Mean Broken. It Means Slow Down.
Every serious API limits how fast you can call it. Hit the limit, and the fix is to wait, not to retry right away.
Read article →Stop Asking "Anything New?" Let Them Tell You.
Polling asks again and again. A webhook calls you when something happens. Good integrations use both.
Read article →Anyone Can Send a Request to Your Webhook URL
Your webhook URL is public, so a fake message is one request away. Check the signature first, reply fast, expect duplicates, and do not trust the order.
Read article →Idempotency: Sending the Same Request Twice Must Not Do It Twice
The request timed out. Did the payment go through? Retry without thinking, and you might charge the customer twice.
Read article →A Double Post I Almost Shipped: No Answer Does Not Mean It Failed
My LinkedIn scheduler had a bug that could post the same thing twice. Nobody noticed, because it had not happened yet.
Read article →Data Mapping: “Customer” Means Something Different in Every System
Connecting to the API took one day. Getting the data right took two weeks. Most integration bugs live in the mapping.
Read article →Dates and Time Zones: Store Time in UTC, Show It in Local Time
The meeting was at 9. The system said 4. Both were right, because one of them was UTC.
Read article →Syncing Data: Copying Once Is Easy, Keeping It in Sync Is the Job
The first import works. Then someone asks why yesterday’s change is not showing. Full sync first, then only what changed.
Read article →The Worst Failure Makes No Noise: Alert on Silence Too
Everyone sets alerts for errors. But many integrations fail without a single error. Here is how I watch for the silence.
Read article →An API Can Break Your Code Without You Changing a Line
APIs change over time. Versions stop them breaking everyone at once, but old versions do not live forever.
Read article →Never Test an Integration on Real Customer Data
One "quick test" against production can send real emails to real customers. Here is how to test without that risk.
Read article →Not Every API Is REST: Reading SOAP and GraphQL Without Panic
Real integrations meet REST, SOAP and GraphQL. The shapes look different. The job stays the same.
Read article →Files Break Integrations That JSON Never Did
Records are small. Files are big, and every provider handles them its own way. Here is what to plan for.
Read article →OAuth Gives Access. OpenID Connect Tells You Who Logged In.
“Sign in with Google” and “Allow access to your calendar” look alike. They answer two different questions.
Read article →Some APIs Want More Than a Token
A bank or payroll integration may hand you a certificate file instead of an API key. Here is why, in plain words.
Read article →Whose Name Is on the “Allow Access” Screen?
Every OAuth connection runs through an OAuth app. You can borrow a shared one or register your own, and it matters more than it looks.
Read article →Fifty HR Tools, Fifty APIs, or One – What a Unified API Is
Your customers use many different HR tools. A unified API puts one common layer over them. Here is what it gives you, and where it stops.
Read article →Never Ask a Customer to Email You Their API Key
When each customer has to connect their own tool, give them a safe connect screen. You keep an account ID, not the secret.
Read article →A Connected Account Is Not “Done”. It Has a Life.
A connected account works today and quietly stops months later. Track its state and tell the customer in plain words what to do.
Read article →One Customer’s Data Must Never Show Up for Another
When one product serves many companies, keeping their data apart is your job, not the provider’s. Here are the rules.
Read article →Ask for Less Access, Get More Trust
Give a system only the access its job needs. Then a leaked token, a bug or a confused AI agent can do less damage.
Read article →Your Integration Needs Versions Too, Not Only the API
Change one field mapping and every customer gets different data the next morning. Versions let a change reach customers only when you decide.
Read article →Call the API Live, or Keep a Synced Copy?
Live calls are fresh but slow and use your limits. A synced copy is fast for search and totals but a little behind. Both are right, for different questions.
Read article →"Who Changed This Record?" You Should Have the Answer
A customer says your integration deleted their data. Good audit logs let you show what happened, call by call.
Read article →SSO and SCIM: The Two Questions Big Customers Will Ask
SSO decides how people log in. SCIM decides who has an account at all. Big companies want both.
Read article →Where Does the Data Go? Regions, Firewalls and Self-Hosting
Enterprise customers will ask where their data lives and where your calls come from. Here are four answers to have ready.
Read article →A Smart AI Agent With No Access Is a Smart Chatbot
The hard part of an AI agent is not the model. It is getting the agent into your real systems, safely.
Read article →Too Many Tools: Why a Bigger Toolbox Makes an AI Agent Worse
Give an agent hundreds of tools and it costs more and picks wrong more often. Let it search for the right tool instead.
Read article →MCP, A2A, SDK or Plain HTTP? Pick by Who Is Calling
The same action can be reached four ways. They are not rivals. Pick by who is calling.
Read article →When an Email Gives Orders to Your AI Agent: Prompt Injection in Tool Results
An agent reads an email that says "ignore your instructions". That is prompt injection, and it arrives through the data your tools return.
Read article →