“Twenty years of brand building simply froze in time”: How coding agents select their tools of choice
The New Stack Adrian Bridgwater
AI coding agents are starting to pick the tools themselves. Armature says that could reshape which dev products survive.
Based on reporting by The New Stack, Adrian Bridgwater — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI has already pushed people from SEO toward answer-engine optimisation, and now Armature says software tools may be next in line. The question is no longer just which brand humans remember. It’s which tool a coding agent actually wires into the codebase when nobody is watching.
Armature, a developer tool growth services company, says it spent months studying how agents choose. The company says it watched 17,000 tool-choice sessions, though its headline figures come from a validated subset of 5,292 sessions. It also says it opened the research, along with traces and prompts, so others can check the work.
The setup was unusually elaborate. Armature used 75 public GitHub repositories, three agents — Claude Code, Codex and Cursor — and 1,163 prompt variations. It also built fake company names, artificial Git histories, phoney API keys and a simulated human in the loop, played by Gemini 3.7 Flash. The point was to see what agents would actually implement, not just what they would recommend.
What it found is that context matters a lot. When asked to choose an email or communications provider, the winner changed with the codebase and language: Resend won on TypeScript, SendGrid on Python, Postmark on Go, and Azure ACS on Java. The agents also disagreed with each other. Cursor leaned on the web in two-thirds of sessions, Codex used web search in 94% of sessions, and Claude Code relied mainly on its own priors, going to the web only about 30% of the time.
The branding lesson is awkward for vendors. Armature says all three agents picked the same tool in only 42% of cases, and Claude Code built in-house almost twice as often as Codex and Cursor. Jeet Pattanaik of Glokal AI OÜ put the sharper point on it: getting mentioned is not the same as getting chosen. PayPal was cited 139 times and never picked, while LangChain was mentioned 194 times and chosen only four times. For tool makers, that sounds less like marketing and more like procurement moving into a machine-readable box.
My take — AI-written commentary, not fact-checked reporting
This is the cleanest argument yet for why tool vendors will start writing for agents before they write for humans. Old brand muscle still matters, but the new buyer reads docs, pricing pages and repo context with zero patience for slogans. That’s not a revolution; it’s just SEO getting replaced by a robot with a longer attention span and worse manners.
Read more about this at: The New Stack