Coding agents were built for engineers. Turns out they work for everyone else, too.
Coding agents like Antigravity write code, run it, read the error, fix it, and try again. They keep going until it works.
For software engineers, the value was obvious from day one. More code, shipped faster, with less grind.
That was the expected part.
The loop that fixes code works on any job that runs on data. That's nearly every job.
Take Maria. She manages a roofing distribution branch in McKinney, Texas. She doesn't write code and has no interest in learning.
But the morning after a hailstorm, her job is a data problem. Storm reports, the CRM, the ERP, an inbox full of orders, and voice memos from contractors standing on damaged roofs. Pulling it together used to take her whole morning.
She never sees the code. She sees the answer.
Maria asks three questions in plain English. For each one, the agent quietly writes the code to answer it.
It joins hail reports to her contractors by ZIP code and ranks them by damage and open pipeline.
It reads roof photos, listens to voice memos, and parses emails in English and Spanish to build one order book.
It queries the ERP, finds the shortfall, and plans transfers from sister branches.
A chatbot can tell you what to do. An agent that calls fixed tools stops the moment one of them breaks.
And enterprise systems break all the time. An ERP upgrade renames a field, or a unit switches from bundles to squares.
A coding agent reads the error, checks the new schema, rewrites its own tool, and runs it again. Maria never files a ticket. She never even knows it happened.
Agents used to call tools. Now they write them.
Maria asks for a storm-response page: the hail map, ranked contractors, stock against demand, and replies drafted in each contractor's language.
No vendor sells that, and IT would never prioritize it. The agent builds it in minutes. When software costs almost nothing, a single morning is reason enough to build it.
Nothing goes out until she approves it. Then the emails send, the transfers dispatch, and the CRM logs the follow-ups.
6:00 to 6:12. That used to be Maria's whole morning.
That's the agent assisting. Next it acts, and then it anticipates. Prompts become triggers, and people review the results instead of running every step.
Same branch, same storm, 72 hours earlier. At 2:00 AM on Friday, a nightly agent sees a 68% chance of large hail over McKinney. It moves stock in from sister branches, pre-clears contractor credit, and drafts customer alerts.
Nobody wakes up. If a system changed overnight, it fixes its own tools and keeps going. Then it stops at the edge of its authority.
At 6:00 AM Saturday, Maria gets one page.
Already handled. $23,630 of a $25,000 cap.
Sent to a person, with the agent's recommendation and the data behind it.
Every action can be reversed until the trucks roll.
Logged, hashed, and explainable after the fact.
Where the agent stops is a business decision, not a technical one. When I put one of those escalations to a room of executives, 32 of 50 approved it and the rest held. Both answers are reasonable, and that's why the escalation exists.
On Monday, Maria's branch is the only one with stock. Her contractors win the jobs.
Coding agents didn't stay with engineering. They're turning up in branches, warehouses, and back offices, wherever someone has data and a deadline.
The tool your team needs this morning gets built this morning, by the person who needs it.
Value creation moves from software projects to agents in every function. Here's the plan.
Your data used to be the moat. Now it's how fast your people turn data into action. Soon it's how fast your agents do it without them.