Industry

AI-generated text

How AI Agents Are Changing the Role of Software Engineering Expertise

As AI agents take on more coding, developers’ expertise is shifting toward designing workflows, setting safeguards and assessing results.

How AI Agents Are Changing the Role of Software Engineering Expertise

Artificial intelligence agents are writing an increasing share of program code, changing developers’ daily work and the role of their expertise. Louise Cox argues that this expertise is not becoming obsolete: it is instead expressed in how developers organize agents’ work, set boundaries and decide what evidence is needed before accepting the results.

Gergely Orosz recently wrote that modern software engineering no longer happens without large language models and will not do so in the future. Some developers miss writing code by hand, while others question how much agent-generated code they should read. As Kent Beck has observed, writing and reading code also gave developers a sense of competence.

Cox recalls that, ten years ago, the tangible act of creation and the correction of mistakes during live coding demonstrations at conferences were particularly compelling to her. She considers it understandable that developers feel a sense of loss as agents increasingly take over this hands-on part of their craft.

Expertise moves to a higher level of abstraction

In a recent article, Tim O’Reilly argued that O’Reilly’s business is built on cultivating and sharing human expertise, making it important to find the areas where that expertise still matters. Cox maintains that experience gained in building and maintaining software now helps developers direct and assess agents’ work.

Agents can amplify the guidance they receive, but developers must first understand the task and the code. Avi Flombaum, founder of Flatiron School, told Cox that developers’ value was never limited to writing or reading code. Experience remains important for understanding complex problem domains, defining what needs to be built and applying engineering judgment.

In a separate conversation, Ken Kousen noted that developers have long worked with code produced by third parties: previously, it was called outsourcing. Designing AI workflows, setting guardrails and managing agents therefore offers another way to apply accumulated engineering knowledge.

The whole development process must change

In so-called software factories, companies are employing agents across multiple stages of the development lifecycle. The aim is to take work from an initial signal through implementation, validation and deployment without requiring a person at every step.

Faster code generation alone, however, does not solve problems elsewhere in the process. Without other changes, teams can spend many tokens creating pull requests that still need review. Anthropic’s playbook for an AI-native software development lifecycle says faster implementation places greater pressure on planning, testing, review, handoffs and approvals.

People must therefore design software-factory workflows, define the conditions for work to proceed and decide which exceptions require human attention. Anthropic reports that Claude authors about 80% of the code merged into its codebase, with a review process combining automated checks and human oversight according to risk.

The cost of code quality and context

The state of an existing codebase also affects agent efficiency. In a study of C++, Java and Python codebases, Adam Tornhill and the CodeScene research team found that agents working on unhealthy code used 35–45% more tokens to complete the same tasks. Refactoring the parts of a codebase agents frequently touch may therefore be worthwhile. Humans still need to identify the important areas, decide what a better solution looks like and confirm that behavior has not changed.

Appropriate context can also reduce costs. In Port’s testing, providing structured context cut costs by about 58% compared with giving agents direct access to separate MCP servers. When agents also received instructions on where to find relevant information, savings rose to around 80%. These results are specific to Port’s experiment, and the work behind them requires knowledge of important relationships within a system and keeping that knowledge current.

Evaluating the workflow, not just the change

Testing can show whether an individual change works, while workflow evaluation asks how reliably the process produces acceptable changes and when a person needs to intervene.

In Evals for AI Engineers, Shreya Shankar and Hamel Husain recommend starting with traces of what a system actually did. In the first stage, called open coding, evaluators read those traces and make brief notes about failures. During axial coding, they group the observed problems into failure categories. Experience debugging production systems can help developers identify more quickly which failures matter.

Cox argues that engineering expertise now also appears in the problems developers define, the standards they set and the evidence they require before a change ships. She stands by her December 2025 assessment that AI rewards strong engineering rather than replacing it. Ben O’Mahony has observed that people with extensive experience in agentic engineering have moved beyond the initial sense of magic and returned to treating the work as engineering.

On the same theme, Avi Flombaum is launching a new series, The Agentic Rethink, on the O’Reilly platform on Wednesday, January 13, 2027.