tl;dr
I’m building a Workday Extend app in public, off-hours, with Claude Code and Workday’s Developer Agent doing most of the typing.
Getting from a bare laptop to a deployed hello-world page, with the agent able to edit it, took 40 commits and 25 logged surprises.
Almost none of it was code. It was finding out how the platform behaves on a personal developer account.
The docs describe a customer tenant. A personal account is a different product, and the differences hide in flags, gateways and missing checkboxes.
“Install the extension” never means install the extension. Java will find you. It always finds you.
WDCLI will print
Success: falseand exit zero. Read the output, never the exit code.Four sign-ins, four clocks, none of them refresh. You are the keep-alive.
With a map, this is two evenings. I wrote the map.
“Hello, Triage Console.”
That’s the whole page. Three words and a comma, rendered in a Workday dev tenant at 9:00 on a Thursday morning.
It took forty commits to get there. Forty. For a comma.
What I’m building, and why in public
Integration history ages out. The Insights API keeps fourteen days of events, and the institutional memory of why an integration broke lives in somebody’s inbox until that somebody leaves.
So I’m building an Extend app that keeps failure history inside Workday, groups repeat failures, routes them to an owner, and records why each one broke and what fixed it. It runs in a free Workday Developer Program tenant, with fictitious data, on my own laptop and my own time. Claude Code wrote most of the files. Developer Agent wrote some. I clicked Save, and did a bunch of tenant work (Save gets its own issue. Save has earned its own issue.)
And I started out confident, which is the important part of this story. A fresh-ish laptop. Nothing installed. No legacy junk, no mystery PATH entries, no ghosts of SDKs past. A clean slate.
I genuinely believed a clean machine meant a clean run. Reader, a clean machine just means every problem gets to introduce itself personally.
Start with the terms, not the tools
Before installing anything, I read the program terms. Articles about the program need Workday’s written consent, so I asked before drafting a word of this (Thanks Nick!). And a personal tenant is permanently deleted after seven days without a login. Workday does email you before it happens, and one login resets the clock. Still. It’s a houseplant that files for divorce if you don’t text it twice a week.
So: login reminders, Sunday and Wednesday.
The first dependency of this whole project arrived before a single runtime: a recurring calendar reminder.
A personal account is a different product
I requested the tenant at 7:18 PM on a Wednesday, and it was live a few minutes later. Within the hour, the docs and the tenant had disagreed three times.
The Insights Events Enabled checkbox the docs describe? Not there. Just a blank space where a feature used to be in somebody else’s tenant.
wdcli auth login sent my personal account to company SSO, where it waited, politely, until it timed out. The real command is wdcli auth login -e personal. And here’s my favorite part: when you’re not logged in, the CLI’s own error message tells you to run the wrong one. It’s a GPS that notices you’re lost and confidently routes you into the lake.
It was a little bit my fault. The right flag is sitting in wdcli auth login --help. I did not run --help. I had a fresh laptop and confidence. The doc pages don’t help: one says wdcli auth login, another says the older wdcli auth:login, and neither works for a personal account.
The API examples I copied pointed at the US gateway, which returns a 401 for personal accounts. The right one is listed in a reference table, and on my own Developer Site profile page. I found both after the 401. And on that gateway, the Insights API, the API my entire design was built around, returns a 502.
So the foundation of my app didn’t exist. Cool. Cool cool cool.
By 8:18 PM I had a WQL query returning real integration events instead. Not fourteen days of them. History back to April 2023, inherited from the template tenant. Losing the headline API turned out to be the best thing that happened all week, which is not a sentence I expected to type about a 502.
The docs describe the tenant you’d buy. You’re standing in the one you were given.
“Install the extension” never means install the extension
First, in fairness: you don’t have to do any of this. Workday’s own Developer Agent tutorial runs entirely in the browser, with zero installs. I wanted the agent on local files, in git, with Claude Code driving. This section is the price of that choice.
Developer Agent for Claude Code takes four pieces: a VS Code extension, a Claude Code plugin, WDCLI, and a sign-in inside a VS Code panel that stays asleep until you click on it. Like a cat. A cat that holds your deployment pipeline.
And then the Workday Toolkit’s live IDE validator. Which is a Java program.
Of course it’s Java. It’s always Java. I’ve worked in enterprise software for twenty years. I have seen things. And somehow I sat down at a fresh laptop thinking the enterprise would leave Java at home this time.
Without Java, the extension logs “activated successfully” and validates nothing. It’s a phantom smoke detector with no battery, cheerfully blinking green. So I installed Java, set JAVA_HOME like a responsible adult, and it still couldn’t find it. On Windows, the extension looks for bin\java. The file is called bin\java.exe. So JAVA_HOME can never work, and Java has to be on your PATH. I know this because I read the extension’s source code. On a Thursday. On purpose.
Then I ran wdcli tenant register-default-agent, which sounded exactly like the setup step. It failed with no reason given and convinced me my account wasn’t entitled to Developer Agent. It was. Different feature, similar name.
An opaque error invites the wrong conclusion, and you’ll volunteer for it.
Hello world is assembled by hand
wdcli app create sounds like it creates an app. It writes one file, appManifest.json, and nothing in it can deploy. It’s ordering a house kit and receiving the deed. Congratulations on your new home. Please assemble the house.
So I assembled the house by hand: five files copied from Workday’s sample catalog, with an application ID that has to match across the manifest, the AMD and the SMD.
Then the first deploy printed Success: false and exited with code zero. So the deploy failed, said so in words, and exited zero anyway. Claude Code reads exit codes, so that zero tells any script or agent listening that a failed deploy succeeded. It’s a waiter telling you “enjoy!” while your plate is on fire.
Fix, upload, deploy. 9:00 AM Thursday: “Hello, Triage Console.” Sixteen surprises in.
Then the agent, once I opened the right folder
The last stretch was getting Developer Agent to edit that page.
It couldn’t, at first. My app lived in a subfolder of the repo, and the agent treats whatever folder VS Code has open as the app, so every single write landed one directory short. An hour. The fix was opening a different folder. The identical request then worked on the first try, which is the most humbling possible way for a problem to end.
On the way, one delegated turn hung silently for thirty minutes, waiting on a VS Code window that had already closed.
At 11:46 the agent edited the page, I validated, uploaded and deployed, and the tenant answered success: true. Twenty-five surprises. Functional. Humbled, but functional.
Two straight complaints
The people who build this will read it, so I’ll be plain.
The clocks are too short. A personal build runs on four separate sign-ins: the Developer Site account, the tenant token, the tenant in the browser, and the Developer Agent panel. Each keeps its own clock. None of them refreshes. The account token has a hard ceiling of about two hours and ten minutes, the tenant token underneath it expires every sixty minutes, and the browser and VS Code panel keep their own idle clocks. In practice, something wanted me back roughly every ninety minutes. It felt like half my time went to scripting around timeouts or clicking the button that lets me back into the tenant. I wrote a login script that renews early and warns before the hard limit. It works. I’m still the keep-alive.
There’s still a lot of human in the loop. The last time I did Extend work, it was called Workday Cloud Platform. I know enough to be dangerous, and I know an app with an agent, a data model and orchestrations isn’t a four-hour job. I’m genuinely impressed: four weeks of planned data model and orchestration work landed in three days. But the speed came from the AI side. On the Workday side, a person still signs in, clicks Save, copies sample files by hand, and waits for an agent to activate.
The machines did four weeks of work in three days. They spent a lot of it waiting on me.
Where Workday deserves credit
The program is free, and anyone can join. The docs are published as Markdown with an llms.txt index built for AI tools, which is more forward-looking than most vendors manage. Developer Agent officially supports Claude Code; the datasheet said Extend Professional only, and it was wrong in my favor. And the sample catalog answered more of my questions than any agent did.
And in fairness to both complaints: this is a beta program for individuals. A customer tenant on a paid tier may never hit half of this.
The two-evening version
Every trap above is in a companion build guide, in order, with the command that gets you past it. Call it two evenings once you factor in account approval, tenant provisioning, and four and a half to seven hours of hands-on setup.
And bookmark Workday’s own front doors before you start. Quick starts live at developer.workday.com/documentation, videos at youtube.com/@WorkdayDevelopers, and the blog at dev.to/workdaydevelopers. The Developer Forum (forum.developer.workday.com) is watched by Developer Relations, Product and Engineering, so ask there first. Odds are, if it broke for you, it broke for somebody there last week. Urgent? Email devprogram@workday.com. Built something? Share it on the Dev Program GitHub at github.com/Workday/Developer-Relations.
Coming soon: the app itself, and the first time it recognized a failure it had already seen.
Hello world isn’t the first line of your app.
It’s the receipt for everything the docs left out.
— Mike
Director HR Tech | Java’s Latest Victim
P.S. In 3D printing, the hello world is a Benchy: a little tugboat everybody prints first to prove the machine works. Justin’s teacher gave him a well-used (aka - broken) Ender 3. He nursed it back to health with a new hotend and a new build plate. Every part is right. It still will not print a Benchy. I didn’t have the heart to tell him that a few days later I built my first Workday agent, got every part right, and then spent Friday to Sunday night unable to get it to activate. Sometimes a hello world only proves you haven’t found the last broken thing yet.
P.P.S. You made it through all that techy-speak. Here's your prize: one Markdown file that does most of this setup for you. Drop it in an empty folder, start Claude Code there, and tell it to walk you through. Claude does the installs, the checks and the traps. You do the passwords and the clicking, which is the part no agent should touch anyway. It's the map, except now the map drives. Fair warning: it's fresh-ish too. If it finds a new broken thing, tell me.
The Department of First Things First. For the people who do the work.



