A study published last week by Armature, a developer tool growth services company, found that AI coding agents are now the ones choosing which software tools get built into codebases — and that twenty years of vendor brand-building has been "frozen in time" as agents default to legacy reputations while missing current capabilities. The company analyzed 5,292 validated tool choice sessions across three coding agents (Claude Code, Codex, and Cursor) to understand how these systems discover, evaluate, and select developer tools. According to co-founder Theodore Otzenberger, agents "reach for Docker the second containers come up, then draw a blank on the sandboxes it offers now, so that it actually ends up not picking the tool."

The experimental analysis examined 1,163 prompt variations across 75 distinct codebases, simulating different developer personas ranging from vibe-coders to senior engineers at enterprises. The research found that all three agents selected the same tool in only 42% of cases, and Claude Code built solutions in-house nearly twice as often as Codex and Cursor (19% versus 10%). Repository context proved decisive: when asked to choose an email service provider, four codebases in four languages returned four different winners — Resend prevailed on TypeScript code (55 out of 89 runs), SendGrid on Python (22 out of 24), Postmark on Go (20 out of 24), and Azure ACS on Java (22 out of 23). The agents also disagreed on where to look for information: Cursor based decisions on the web in two-thirds of sessions, Codex used web search 94% of the time but relied on operators like "site" in nine queries out of ten, and Claude Code searched the web only about 30% of the time but browsed three times more pages than Codex when it did.

Armature's study revealed a striking disconnect between brand visibility and actual selection. PayPal was cited 139 times but never chosen. LangChain was the most-mentioned framework at 194 citations, yet it was picked only four times. The company's co-founders, Otzenberger and Louis Scremin, ran the full analysis — later expanded to seventeen thousand total tool choice sessions — using public GitHub repositories, then debiased the statistics to match an ideal panel distribution based on publicly available data. To guard against agent bias, they used fake company names, artificial Git histories, and phony API keys, with a simulated human orchestrator played by Gemini 3.7 Flash. According to the report, adding this simulated human "in the loop reduced the leader [tools] and cloud platform-native solutions dominance towards a more realistic picture."

The research suggests vendors face what Otzenberger calls a "life-or-death impact" as agents replace human developers in tool procurement decisions. The report explains that a tool's reputation from past successes follows it into the model weights, but that legacy recognition now operates "under a different kind of gravity" — agents need tools mentioned, elevated, and presented as immediately usable or those tools "simply stop existing in the stack" tomorrow. What tips an agent's choice is often surprisingly small, according to Jeet Pattanaik, founder of Glokal AI, who noted that "the lever isn't brand awareness any more, it's whatever the agent happens to read at the moment it decides." Vendors will likely optimize documentation and pricing pages next, he said, presenting information "for a non-human reader that doesn't skim, isn't charmed by a logo, and takes a retention footnote completely literally." Otzenberger said the company published every trace and prompt so anyone can verify the findings and see where they stand, adding that "the decision makers are changing and it's now an engineering problem to understand them." The team plans to rerun the full study on newer agents and models soon. If agents continue delegating entire tasks end-to-end as models improve, the shift from human to machine procurement may rewrite how developer tools compete for adoption. Vendors who treat agent decision-making as a communications challenge rather than an engineering problem risk losing visibility in the stacks where their products once dominated by default.