Gwen Davis believes that as the use of artificial intelligence agents expands, developers’ value will shift further from writing every line of code toward directing tools, reviewing their outputs, and making technical decisions. The article suggests three practices: managing agents, not accepting the first answer, and investing the time saved in solving broader problems.
Developers’ work is changing as artificial intelligence tools move from helping write code to carrying out larger portions of software tasks. According to Gwen Davis, writing code alone is no longer sufficient; defining the problem, providing the appropriate context, evaluating outputs, and explaining technical trade-offs before adopting a solution are becoming increasingly important.
1. Direct Artificial Intelligence Instead of Simply Using It
The article explains that execution in an artificial-intelligence-based environment begins with clearly defining the work, followed by distributing tasks and reviewing results. For a task to add an authentication route, for example, one agent may handle preparing the implementation, while another prepares the documentation and a third sets up the test suite.
This does not eliminate the developer’s responsibility for the final result. The practical change is that the developer spends less time manually implementing every part and more time defining what is required, coordinating the agents’ outputs, and deciding what is suitable for review or release. Davis concludes that learning to direct artificial intelligence agents has become an independent practical skill.
2. Do Not Trust the First Answer Without Reviewing It
Tools may produce a seemingly convincing solution within seconds, but it may contain flaws that do not appear during a quick read. The article uses the example of an SQL query that returns the latest order for each customer: the solution may overlook handling identical timestamps, fail to suggest an appropriate index, or suffer performance degradation with large tables.
Davis suggests using a second model to critique the first model’s work and then applying engineering judgment to both responses. She notes that models differ in their strengths and blind spots, and mentions that the Rubber Duck agent built into GitHub Copilot uses a second model to critique plans, code, and tests. However, human review remains necessary; the idea is not to replace human judgment, but to add a critical perspective before proceeding.
3. Use the Time Saved to Solve Bigger Problems
When artificial intelligence handles more of the implementation, developers can direct the available time toward understanding customer needs, evaluating architectural trade-offs, designing systems, and making decisions that the tool cannot make on their behalf.
In an example involving the addition of dark mode, artificial intelligence may implement the changes, generate the tests, and update the documentation. The developer’s checklist, however, includes verifying the problem customers are experiencing, reviewing the architectural trade-offs, checking accessibility, defining success metrics, and approving the solution.
What Changes in Practice?
The central message is not that programming skill has lost its value, but that the scope of responsibility is expanding. The developer remains responsible for the quality of the result, but needs to combine the ability to direct tools with the ability to detect their errors and connect implementation to the project’s actual goal. The article offers general practical guidance, but does not specify metrics for measuring the impact of these practices or limits on the tasks that should be delegated to agents; therefore, implementation decisions remain tied to the nature of the project and the level of risk.