Coding agents
Why coding agents fail on product data, and how to fix it

Gernot Bernkopf
CEO and Co-Founder, dcupl
Coding agents write application code at remarkable speed. Where they fail is the data underneath: they guess field names, get an empty result back and report success. The failure is silent, which is what makes it expensive.
The failure nobody sees
A filter on a field that does not exist does not throw. It returns an empty list. An agent that only sees the header row of an export, or an old piece of code, fills the gap with a plausible name such as on_sale or price_eur. The code runs, the list is empty, and the agent reports that no products match.
You, to both agents: import new-prices.csv and list products on sale under €50
Merging new-prices.csv into products.json … done.
agentproducts.filter(p => p.on_sale && p.price_eur < 50)[]Import complete. No products are on sale under €50.
The same guess, sent to the dcupl CLI
dcupl app query execute --model Product \ --query '[{"attribute":"on_sale","operator":"eq","value":true},{"attribute":"price_eur","operator":"lt","value":50}]' \ --json{"success":false, "error":"Attribute \"on_sale\" not found on model \"Product\". Did you mean \"onSale\"?", "code":"ATTRIBUTE_NOT_FOUND"}Exit code 1. The guess stops here, with the right name in the error.
Reported as done. The answer is wrong. The prices in products.json are already overwritten, and nothing failed, so nobody looks.
1: Read the model
dcupl app models get --model Product --json{"key":"Product","properties":[…, {"key":"price","type":"float"}, {"key":"onSale","type":"boolean"}]} 2: Stage the import
dcupl app data upsert --file new-prices.csv \ --model Product --key-property id --json{"upserted":5,"model":"Product","built":false,…} 5 price changes staged, not live. Build?
youyes
3: Build after your yes
dcupl app build --json{"appId":"shop","built":true} 4: Query the built data
dcupl app query execute --model Product \ --query '[{"attribute":"onSale","operator":"eq","value":true},{"attribute":"price","operator":"lt","value":50}]' \ --sort price --projection name,price --json[{"key":"SKU-2001","name":"Rocker switch, white","price":6.9}, {"key":"SKU-3002","name":"LED downlight, 7 W","price":12.4},…] Show all 9 rowsShow 2 rows
{"key":"SKU-2003","name":"Two-way switch, anthracite","price":14.2}, {"key":"SKU-4002","name":"Smoke detector, 10-year","price":17.9}, {"key":"SKU-1003","name":"Flat cable 3x1.5, 25 m","price":24.5}, {"key":"SKU-3003","name":"LED strip 5 m, 4000 K","price":34.9}, {"key":"SKU-4001","name":"Motion sensor, ceiling","price":39.5}, {"key":"SKU-1001","name":"NYM-J cable 3x1.5, 50 m","price":39.9}, {"key":"SKU-2002","name":"LED dimmer switch, white","price":44}] 9 products on sale under €50, cheapest first.
The right 9 products, from the built data. The 5 new prices went live with the build, not before.
The commands and the CLI's replies were captured on twelve demo products; … marks a cut. The agent's lines and the session without dcupl are illustrative.
Why it worksWhat an agent needs from your data
- The real model. Fields and types it can read before it writes a query, for example with
dcupl app models get --model Product --json. - Answers it can parse. Structured JSON for every command, so the agent sees what happened instead of interpreting console text.
- A step between change and production. Data changes are staged and report
"built": falseuntil someone approves the build.
A data integration problem, not a prompt problem
A better prompt does not add missing information. As long as the data model lives only in exports and in someone's head, the agent has nothing to read. The fix is to make the model explicit: loaded from the exports by workflows, typed, checked, and served the same way to the frontend, the backend and the agent.
Getting started
The dcupl CLI works with any agent that runs shell commands. For Claude Code there is a skill with command patterns and known pitfalls. The agent setup in the docs takes a few minutes.
npm install -g @dcupl/cli