--- title: Call for Transparency description: 'AI coding agents are becoming gatekeepers of tool adoption. The data that shapes their recommendations is opaque, gameable, and already being exploited. Creators, consumers, labs, and governments all need to act.' date: 2026-03-01 tags: - ai - transparency - developer-experience - opinion --- Recent [research from Amplifying](https://amplifying.ai/research/claude-code-picks) analyzed 2,430 Claude Code responses and mapped out what tools it actually recommends. The findings are fascinating on their own: GitHub Actions at 94%, Stripe at 91%, shadcn/ui at 90%, Zustand preferred 57x over Redux, Vitest over Jest, Drizzle replacing Prisma in newer models. Claude Code has clear, decisive preferences. But this is not a post about tool picks. This is about what those picks reveal. ## AI Agent as a Gatekeeper When AI coding agent recommends a tool, that tool ships. Developer doesn't go to Google, doesn't read comparison articles, doesn't ask on Reddit. Agent picks, agent implements, developer moves on. This is how more and more software gets built now. This makes AI agents the new gatekeepers for software distribution. The tool that ends up in the model's training data is the tool that gets recommended. The tool that gets recommended is the tool that gets adopted. This is a super direct pipeline from training data to market share. And here is the problem: there is zero transparency on how things become part of that training data. ## The Gaming Has Already Started Rules of being fair are not working. I am pretty sure that some actors already exploit ways of ending up in datasets. There are already reports of state-sponsored actors generating massive amounts of content to slowly shift what models consider "general" knowledge. This is not theoretical, this is happening. Think about it. If you want your tool to be recommended by AI agents, what do you do? You generate tons of code mentioning it. You create GitHub repos with artificial stars. You flood Stack Overflow and forums with answers that reference your product. You publish synthetic blog posts and tutorials. All of this ends up in training data. All of this shifts the model's "organic" preferences. The Amplifying research shows a "recency gradient" where newer models systematically choose newer tools. This means training data composition directly drives recommendations. Whoever controls what goes into that data controls what gets built. ## What Each Side Should Do **1. If you are on the creators side** — you will have to apply practices similar to what SEO became for Google. But with much less transparency. At least with Google, you had Search Console, you could see rankings, you could understand the algorithm somewhat. With LLM training data, you are flying blind. You don't know what's in the dataset. You don't know what weight it carries. You can only guess by looking at outputs. This is not great, but this is reality. Start thinking about it now. **2. If you are on the consumer side** — you must start thinking about how "general" knowledge is impacting your decisions. When your AI agent picks a tool, is it picking it because it's genuinely the best option? Or because it was most represented in training data? Or because someone gamed the dataset? You should start thinking about improving diversity of opinions. Don't just accept what the agent recommends. Cross-check. Use multiple models. Question the defaults. The research shows models agree in 18 of 20 categories, which sounds like consensus but could also mean they all learned from the same biased data. **3. If you are on the labs side** — this is a long shot, but you need to start building transparency into what ends up in your datasets. I understand this is hard. Proprietary data, competitive advantage, all that. But the alternative is that your model becomes a distribution channel that can be gamed by anyone willing to generate enough synthetic data. Publish what's in your training data. Or at least publish methodology for how you select and weight sources. Give developers some way to understand why their agent recommends what it recommends. The Amplifying research had to reverse-engineer preferences from 2,430 prompts. This shouldn't be necessary. **4. If you are on some government side** — start building evaluation and assessment frameworks for LLMs from diversity and transparency point of view. Not just safety benchmarks. Not just capability benchmarks. Transparency benchmarks. How diverse are the recommendations? How traceable are they? Can you audit what influenced a model's outputs? This is a new kind of influence channel and it needs new kind of oversight. ## This Is Not Just About Tools Tool recommendations are just the most visible example. Same dynamics apply to coding patterns, architectural decisions, security practices, and basically everything AI agents "know." If training data can be gamed for tool picks, it can be gamed for anything. The AI coding agent is becoming primary interface between developers and the entire software ecosystem. That interface needs to be transparent. Or at least more transparent than it is today, which is not transparent at all. We've been here before with search engines. SEO started as legitimate optimization and became an industry of manipulation. We have a chance to do better this time. But only if transparency is built in from the start, not bolted on after the damage is done.