If most software code is written by AI, how will the role of engineers change?

Ankit Jain (안킷 제인), co-founder and CEO of developer productivity platform Aviator, recently wrote in an article for The New Stack that the emphasis will shift from writing code directly to building machines that write code.

He said engineers are likely to be divided not into front-end or back-end roles but into product engineers and platform engineers. Product engineers build products customers use, while platform engineers build the tools used by engineers who build those products. The more machines write code, the more demand will grow for platform engineers who build and refine those machines, he said.

Jain said, "Whether you call it harness engineering or loop engineering, the essence is the same. It is building the tools every engineer depends on to write code, in other words platform engineering."

He also described problems that have emerged in various organisations as AI writes more code. In particular, he said friction increases on the back end as code generation speeds up. Each team writes its own prompts, sets its own guardrails and builds separate dashboards to track what agents got wrong. Nothing is shared, he said, and the whole organisation is solving the same problems separately.

Jain also introduced a case involving Vanita Kumar, a consultant at software consulting firm Thoughtworks. Kumar said that when people talked about a "platform team," it used to be enough to have a single team managing deployment infrastructure, but she found while drawing an organisation chart for a client that a separate team is now needed to manage AI agents. Jain said he expects the roles of the 2 teams to eventually be merged into one.

He said as AI produces large volumes of code, controlling, validating and scaling the code generation process becomes the job of platform engineering. Jain said, "Every company that makes software will effectively become a development tools company."

He said the scope covered by harness engineering is broad.

First is deciding which models and frameworks to approve because it directly affects security and costs. Second is managing usage and scaling. What works for 10 users is a completely different issue from what works for hundreds, he said.

Third is permission management. He said an agent should not be given the same access rights as an engineer simply because it works on an engineer's behalf. Fourth is the harness itself. It involves building a feedback loop that sends issues found by security scanners or linters back to an agent rather than to a person.

Jain said, "Anthropic or OpenAI provide the agents that serve as raw materials, the 'engine.' But that engine has no idea what your codebase looks like," and added, "Only the platform team can refine agents to reflect the team's conventions, acceptable risk level and even budget. If you do not do this work deliberately, each team will end up building its own flimsy version, and when problems arise, no one will be accountable."

He added, "How much productivity you can pull out of AI depends not on which model you use, how strict your adoption policy is, or how quickly you chase this month's buzzword, but on whether you treat the people who build tools, platform engineers, as an afterthought."