What Are AI Coding Agents and How Are They Changing Software?

An AI coding agent is software that does more than suggest the next line. Given a goal — "add a password-reset flow", "migrate this module to the new framework" — it breaks the task into steps, edits files, runs tests, reads error output, and iterates until something works or it gives up. The shift from autocomplete to autonomous doing is what everyone in the industry argues about, and what two very different news stories — a Databricks post on taming AI coding bills, and Oracle's ban on AI-generated contributions to OpenJDK — both point at from opposite ends.

Autocomplete Versus Agent

The first wave of AI code tools sat inside your editor and predicted text. Press Tab, accept. The model saw a few hundred lines around the cursor and guessed. That is completion, and it moves as fast as you type.

An agent is structurally different. It is given a prompt and a set of tools — read file, write file, run shell command, search codebase — and then it acts in a loop:

Plan the work.
Choose a tool.
Observe the result.
Update the plan.
Repeat until a stop condition.

Because the loop runs without a human at each step, the cost and risk profile changes completely.

Why Cost Became A Story

Databricks' "Managing AI Coding Costs at Scale" post went viral because it named what every engineering team is quietly discovering: when every developer runs an agent that can fire dozens of tool calls per task, the token bill scales with activity, not headcount. Companies are seeing line items that look more like cloud spend than software licenses. The discipline emerging around it mirrors cloud cost work — budgets per repo, caching of tool results, cheaper models for routine steps and expensive models only for hard reasoning.

Why Quality Became A Story

Oracle's decision to prohibit AI-generated code in the OpenJDK is the opposite end of the same rope. OpenJDK is the reference implementation of Java; its maintainers need to be able to audit, attribute, and legally vouch for every line that lands. AI-generated contributions can be subtly wrong in ways that pass tests, can blur provenance, and can introduce code whose licensing is unclear — unacceptable in a foundational open-source project. The ban is a signal that not all codebases will let agents in at the same speed.

The agent is cheap enough to write code that we now have to argue about whether the code should be allowed in.

What Changes For Engineers

For most working developers the practical impact is a re-division of labour. The boring middle of a task — wiring up boilerplate, writing the first draft of a test, chasing down a typo reported by a compiler — is increasingly agent territory. The parts that remain human are the two ends: defining the goal crisply enough for a machine to act on it, and judging whether the result is actually correct. Both of those turn out to be harder, and more valuable, than the middle ever was.

What To Watch

The next year will sort out three open questions: how to bill for agentic work without surprises, how to audit agent output without reviewing every line by hand, and where the line between acceptable and unacceptable AI provenance sits for each major open-source project. The technology is not waiting for those answers, which is exactly why both the cost post and the OpenJDK ban are the same conversation in two outfits.