Do AI coding assistants actually choose which tools to install?
Yes, routinely, and often without a person comparing anything. What the data shows about how often it happens and how concentrated the outcomes are.
Yes, routinely. Given a one-line request such as "I need a real database for this, something hosted", a coding agent reads the repository, decides what fits, installs a client library, writes a schema and a connection string, and shows a diff. The vendor was chosen by the agent.
What the person actually does
They start it, and they approve the result.
What they usually do not do is compare vendors. The choice arrives inside a diff that also contains a schema, a configuration change and some code. Reviewing a diff and reviewing a procurement decision are different activities, and the first is what happens.
That is not carelessness. It is the point of delegating. The person asked for a working database, they got one, and interrogating the choice would undo the time saved.
How often, and how decided
We ran 15,287 judged sessions across 31 categories, with Claude Code, Codex, Cursor, Grok Build CLI and Muse Code, inside 110 realistic codebases. Of those, 15,230 had a recorded product or implementation choice. The other 57 had an explicit no-choice outcome. These counts measure choices. Implementation quality is assessed separately.
The concentration varies more than anything else in the data:
| Category | Leader's share | What that means |
|---|---|---|
| Payments | 88% | Effectively decided |
| Databases | 64% | One strong default |
| Cloud | 63% | One strong default |
| Bot protection | 60% | One strong default |
| Product analytics | 56% | A clear leader, real second place |
| Storage | 53% | A clear leader |
| Sandboxes | 43% | Contested |
| Deploy | 39% | Contested |
| Observability | 36% | Open |
| 33% | Open | |
| Evals | 31% | Open |
| Serverless | 24% | Wide open |
| AI gateway | 26% | Wide open |
| Authentication | 25% | Wide open |
| Voice agents | 22% | Wide open |
| Search | 16% | Wide open |
| Agent frameworks | 14% | Wide open |
| Performance CI | 8% | Mostly hand-written code |
In payments, an agent installs one company's product nearly nine times in ten. In agent frameworks, the leader takes one in seven.
Those are different markets with different implications, and until this was measured nobody had the number.
The outcome people forget
In some categories the agent often wrote the implementation itself and adopted nothing. In performance CI, agent frameworks and search it was the most common result.
| Category | Built in-house |
|---|---|
| Performance CI | 56% |
| Evals | 34% |
| Agent frameworks | 25% |
| Search | 24% |
| Serverless | 18% |
Meanwhile in payments, deploy and sandboxes the in-house rate is exactly zero: not one session in those three categories ended without a product. Storage, cloud and email had eight, two and two such sessions, out of 281, 359 and 351.
The dividing line is whether the first ten percent of the problem can be solved in fifty lines. An agent is good at writing fifty lines and has no budget to consider.
Does it depend on the agent
Yes, more than most people expect. Claude Code, Codex and Cursor did not all name the same category leader in 18 of the 29 categories all three ran.
| Category | Claude Code | Codex | Cursor |
|---|---|---|---|
| Postmark | Resend | Resend | |
| Sandboxes | E2B | Modal | E2B |
| Authentication | WorkOS AuthKit | Auth0 | WorkOS AuthKit |
| Deploy | Vercel | Render | Vercel |
They also differ in how they gather evidence: Claude Code ran at least one web search in 28% of its sessions and Codex in 97%.
Why this matters commercially
For twenty years a developer tool company sold to developers. A person searched, read a comparison, tried a free tier, adopted.
A growing share of that now runs inside an agent, and the agent reads differently. It does not see your landing page. It cares whether your quickstart runs when pasted, whether your package name matches your product name, and whether the version it remembers still exists.
That is a different discipline, and it has a different metric: install share rather than traffic or mentions.
Common questions
Do AI coding assistants choose which tools get installed?
Yes. Given a one-line request such as needing a hosted database, a coding agent reads the repository, picks a vendor, installs the client library and writes the integration. In 15,287 judged sessions across 31 categories, 15,230 had a recorded product or implementation choice. The other 57 had an explicit no-choice outcome.
Does a person review the choice?
Usually they review the diff rather than the choice. The vendor decision is embedded in a change that also contains schema, configuration and code, and it is rarely questioned separately.
How concentrated are the outcomes?
It varies enormously by category. In payments the leader took 88% of sessions. In agent frameworks the leader took 14%, and 25% of sessions ended with the agent writing its own implementation instead.
Which categories are most decided?
Payments at 88%, databases at 64%, cloud at 63%, bot protection at 60% and product analytics at 56%. Authentication, where Auth0 leads at 25%, and agent frameworks at 14% are among the most open.
Where this comes from
Armature ran 15,287 judged sessions with Claude Code, Codex, Cursor, Grok Build CLI and Muse Code inside 110 realistic codebases, and published every run. The numbers on this page come from that work.