Gwen Davis: Direct AI agents, don't just use them

GitHub's advice arrives as AI agents take on more of the implementation work inside everyday developer workflows. The company's own example shows a traditional authentication task moving from writing code, running tests, and opening a pull request toward coordinating multiple AI agents that each return a finished piece for review. The article points to that shift as the reason a developer's value is moving toward defining the problem, evaluating AI-generated code, and deciding what ships.

The second tip warns against accepting AI's first answer, using a SQL query example where a second model flags duplicate timestamps, a missing index recommendation, and poor performance on large tables. GitHub links that idea to its GitHub Copilot Rubber Duck agent, which already critiques plans, code, and tests with a second model. This suggests GitHub is positioning such review layers as a normal part of its tooling rather than an optional extra.

The third tip points to Issue #4821 about adding dark mode, where AI handles the build, tests, and documentation while the developer validates the customer problem, reviews tradeoffs, checks accessibility, defines success metrics, and approves the solution. The article's bottom line is that the skills that launch careers are evolving. Whether that advice translates into changed hiring and promotion standards is not something the article establishes, so the practical test may be whether development teams actually reward judgment and AI direction over raw implementation speed. Because the piece is written by Gwen Davis, a GitHub content strategist, it reflects the company's view of how its tools fit into that future.