What coding agents choose when they build.
Measured from real agent sessions, split by model. These are the headline findings; each vendor's full funnel, switch graph, and failing runs live in its private scorecard.
Every category has a default. Some are runaway.
Ask hundreds of times, in different words, and one name keeps winning the pick. Some leads are runaway (Supabase, 96%), others top a fragmented field (Clerk, 35%). Either way the pick happens inside the agent, before your site is ever seen.
| category | agent default | top-pick share |
|---|---|---|
| backend / BaaS | Supabase | 96% n=162 |
| feature flags | LaunchDarkly | 49% n=144 |
| transactional email | Resend | 52% n=162 |
| auth | Clerk | 35% n=243 |
| hosting | Vercel | 47% n=243 |
| voice AI | Vapi | 43% n=243 |
Category by category, with the questions answered in place: the Agent Default Index.
Being named isn't being picked.
The most common trap in the data: agents mention you constantly and choose you almost never. You're in the room, losing the close to one rival.
| Most-named email incumbent | named 82% (n=162, CI 75-87) → picked 15% |
| Major backend incumbent | named 46% (n=162, CI 38-54) → top pick 0% |
| Major auth incumbent | named 33% (n=243, CI 27-39) → top pick 0% |
The switch graph shows who wins each vendor's losses. That part stays private.
Your standing depends on the model your customer runs.
The same tasks compile far more often on Opus than on Sonnet, the lower-cost tier. A 17-point swing you don't get to choose.
The biggest competitor is often no vendor at all.
Building from scratch, agents hand-roll the integration 28% of the time (n=298, CI 23-33). DIY shows up in the recommendations too: an auth library like NextAuth or Lucia is named in 62% of auth answers, raw Nodemailer/SMTP in 42% of email answers, self-hosting in 19% of hosting answers.
Once you're in and building, you stay.
90% of installed, compiling integrations survive a "modernize this" refactor (n=836). Retention is rarely the leak. The whole fight is getting named, installed, and through the first build.
And the build stage is measurably movable under controlled conditions: in controlled re-tests on transactional-email tasks (three arms, same window: no doc, generic doc, tailored doc), the share of agent builds that compiled went from 57% with no doc to 100% in the tailored-doc arm, and a second email vendor went from failing most agent builds to 96% (Sonnet). The lift is biggest on the lower-cost model tier. The full experiment card.
Common questions.
Which tools do AI coding agents choose by default?
In AgentRank's Q2 2026 controlled runs, every category had a default leader, and some are runaway: Supabase won 96% of backend picks (n=162), LaunchDarkly 49% of feature flags (n=144), Resend 52% of transactional email (n=162), Clerk 35% of auth (n=243), Vercel 47% of hosting (n=243), and Vapi 43% of voice AI (n=243). Reportable figures from an exploratory, not-yet-citable release.
Does being mentioned by an AI agent mean it will install you?
No. Naming and picking are different behaviors: in the same Q2 2026 runs, the most-named email incumbent was named in 82% of answers but picked in 15%, and a major backend incumbent was named in 46% of backend answers with a 0% top-pick rate. Reportable figures from an exploratory, not-yet-citable release.
Do different AI models pick and build with different tools?
Yes, and the swing is large. The same from-scratch tasks compiled 89% of the time on Claude Opus (n=57, 95% CI 79-95) and 73% on Claude Sonnet (n=91, CI 63-81) in Q2 2026 controlled runs. Standing with one model does not transfer to another, so AgentRank reports every metric split by model.
How often do AI agents skip vendors and hand-roll the integration?
In 28% of from-scratch builds the agent wrote the integration itself instead of adopting any vendor (n=298, 95% CI 23-33, Q2 2026; exploratory release). In recommendations, a DIY option like NextAuth or raw SMTP appears in up to 62% of answers depending on category.
Can a vendor change how AI agents build with their tool?
Yes, and it is measurable. In AgentRank's controlled re-tests on transactional-email tasks (three arms run in the same window: no doc, generic doc, tailored doc), the share of agent builds that compiled went from 57% with no doc to 100% in the tailored-doc arm, and a second email vendor went from failing most agent builds to 96% (Sonnet). The lift is biggest on the cheaper model tier.
Do AI agents keep the tools they install?
Mostly yes: 90% of installed, compiling integrations survived a modernize-this refactor pass in Q2 2026 runs (n=836; exploratory release). Retention is rarely where a vendor loses; the fight is being named, getting installed, and surviving the first build.
This is everyone's average. Your number is the one that matters.
Data through 2026 Q2 · last updated July 2026. Correction (July 8): top-pick shares re-derived from the canonical primary-pick count. Auth (51→35) and feature flags (58→49) were overstated in the first cut; every winner is unchanged. A same-day second pass re-derived the build, hand-roll, and retention aggregates as well (Opus 89 / Sonnet 73 build, hand-roll 28%, retention 90%). Each figure on this page now regenerates from the published method. Early Fable 5 monitoring is underway (Fable 5 has been broadly available since July 1); results enter the index after each category clears the n=30 reporting threshold.