AI coding agents are reshaping software vendor strategy—brand loyalty no longer guarantees adoption

When AI coding agents like Claude Code and Cursor make decisions about which debugging tools, databases, or package managers to incorporate into software projects, traditional brand recognition counts for almost nothing. Developer tool vendor Armature conducted an extensive analysis of how these agents select their tools, uncovering findings that upend conventional marketing wisdom.

The collapse of brand power in agent decision-making#

Armature analyzed over 17,000 tool-selection sessions to understand how coding agents discover and pick from available options. The central insight is stark: decades of brand establishment have become largely irrelevant in this new landscape. As Armature co-founder Theodore Otzenberger explains, "Watching seventeen thousand tool choice sessions in the analysis undertaken, we saw twenty years of brand building carried out by tool vendors simply frozen in time."

The problem is especially acute for established tools. When Docker emerges as a conversation topic, agents immediately reach for it by default—but then fail to recognize Docker's sandboxing capabilities and may not select it at all. Reputation and brand weight still carry influence, but they function under entirely different rules when evaluated by machine learning models rather than human judgment.

How the study worked#

The research team examined three major coding agents (Claude Code, Codex, and Cursor) across 75 distinct repositories and over 1,100 different prompt variations. The study was performed on public GitHub codebases, with researchers extracting details about programming languages, frameworks, third-party services, and deployment platforms. To prevent bias, the team used synthetic company names, artificial Git histories, and fake API keys.

Crucially, they introduced a simulated human participant into the loop—powered by Gemini 3.7 Flash—who would either accept the agent's top recommendation or ask it to select and implement the best option. This adjustment proved significant: without a human decision-maker, agents showed a much stronger tendency to build solutions from scratch rather than integrate third-party tools.

What context matters most#

Repository context dramatically influences an agent's tool selection. When tested on email service providers across different programming languages, results varied sharply:

• TypeScript codebases preferred (55 out of 89 runs)