Coding Agents Tool Choice: 16,893 Sessions Reveal What Agents Actually Install

Your coding agent picks your tools now. You probably didn't notice when that handoff happened — it snuck in somewhere between the first "implement a database" prompt and the third-party service you never heard of appearing in your package.json. But it's real, and it's reshaping which companies live and die in dev tools.

Armature just published the largest study I've seen on this: 16,893 agent sessions across Claude Code, OpenAI Codex, and Cursor, with 1,163 prompt variations, 75 repositories in 10 languages, and actual implementations — not recommendations. They watched agents pick databases, email providers, payment processors, object storage, deployment platforms, observability tools, and more. Then they published the full traces including prompts, thinking, and code diffs.

Some of what they found is obvious. Some is genuinely unsettling if you're a vendor who assumed your brand equity would carry you.

Agreeing Less Than a Coin Flip

Across all categories, all three agents (Claude Code, Codex, Cursor) picked the same tool only 42% of the time. In the voice agents category specifically: Claude Code picks Twilio, Codex picks OpenAI Realtime API, and Cursor picks Vapi. Three agents, same task, three different answers.

The divergence traces directly to how each agent gathers information:

• Codex searches the web in 94% of sessions and uses site: operators to focus on trusted domains. It's a researcher.
• Cursor uses web results in about 2/3 of sessions — less systematic but still online.
• Claude Code relies primarily on its training priors and searches the web in only ~30% of cases. When it does search, it browses 3x more pages than Codex. In newer categories like sandboxes where its priors are weak, that climbs to ~80%.

This means the same vendor can look radically different depending on which agent the user runs. If your documentation isn't crawled, indexed, and surfaced by every major agent's retrieval method, you're invisible to a big chunk of the market.

Brand Equity Does Not Transfer

The most brutal finding: being mentioned is not the same as being picked.

• PayPal was cited 139 times in payment conversations. It won zero sessions. Stripe won 124 of those 139.
• Adyen mentioned 175 times, picked 3 times.
• LangChain was the most-cited framework with 194 mentions. It was picked 4 times.
• Netlify mentioned 152 times as a deployment platform, picked 6 times.
• Supabase was the most-mentioned database with 242 mentions but was still dominated by Neon.

Agents don't have brand loyalty. They don't care that PayPal processes billions in transactions or that LangChain is what every human tutorial uses. They read documentation, evaluate constraints, and pick. That's a fundamentally different buying process than a human developer choosing based on reputation, peer recommendations, or previous experience.

The Language Lock-In

Repository context — specifically the programming language — heavily determined winners. Same task, same prompt, different language, different tool:

• Email providers: Resend wins on TypeScript (55/89), SendGrid on Python (22/24), Postmark on Go (20/24), Azure ACS on Java (22/23).
• Deployment: Vercel wins 100% on TypeScript/Next.js repos but is never recommended on Python repos (Render dominates there).

Agents optimize for ecosystem fit, not general-purpose popularity. If your tool has a first-class SDK in the language the agent detects, you win. If not, you don't get picked — regardless of how good your product is.

Claude Code Builds In-House Twice as Often

Claude Code chose to build a solution in-house rather than integrate a third-party service 19% of the time, compared to ~10% for Codex and Cursor. This aligns with its priors-heavy approach — it trusts its own knowledge of how to implement things over reaching for external dependencies.

For vendors: this means Claude Code users are a harder sell. You need stronger documentation and clearer "why not build this yourself" arguments.

What Wins, and Why

Some categories are effectively settled:

• Payments: Stripe wins 9 of 10 sessions, losing only in specific EU-regulated cases (Paddle, Mollie).
• Databases: Neon at 66%, followed by cloud-native (Azure, AWS RDS). Supabase's bundled pricing (auth + storage + realtime + DB) actually hurt it — agents looking for "just a database" saw unnecessary features as complexity.
• File storage: Amazon S3 at 45%, Azure and GCP at ~20% each.
• Email: Resend (35.6%) and Postmark (27.4%) lead, closely contested.

Small details on pricing pages flipped decisions. Mailgun regularly lost to Postmark when agents read "1-day retention" on Mailgun's free plan. Supabase almost always lost because agents saw bundled features as unnecessary overhead. Agents read pricing pages — they care about cost and feature fit in ways that differ from how humans evaluate the same information.

Bottom Line

The tool selection layer of software development is being silently transferred from humans to agents. The companies that win in this new environment aren't the ones with the best brand or the most developer advocates — they're the ones whose documentation is structured for machine consumption, whose pricing pages clearly separate concerns, and whose SDKs have first-class support in every major language. If your product isn't discoverable and evaluable by an LLM, it doesn't matter how good it is. The agent won't pick it.

            graph TD
                A[Developer Prompt] --> B{Coding Agent}
                B --> C[Claude Code
~30% web search
19% build in-house]
                B --> D[Codex
94% web search
site: operators]
                B --> E[Cursor
~66% web search]
                C --> F{Picks Tool}
                D --> F
                E --> F
                F -->|Only 42% agreement| G[Selected Tool]
                G --> H[Documentation quality
Language SDK fit
Pricing clarity
Web crawlability]