AI Coding Agents

AI coding agents are software tools that use a large language model to carry out programming work with a degree of autonomy: they read a repository, plan a change, edit files, run commands and tests, and iterate on the results. Claude Code, Codex, Cursor and GitHub Copilot are the most widely used examples. Because these agents also choose the libraries, SDKs and hosted services a project depends on, they have become a distribution channel for developer tools in their own right, one in which the "buyer" evaluating a product is often not a person.

Adoption

Use is close to universal among professionals. JetBrains' Developer Ecosystem research (August 2026, more than 15,000 professional developers surveyed May to July) found 90% using AI coding agents at work at least weekly and 68% daily. Claude Code was used at work by about 39% of respondents, GitHub Copilot by 21% (down from 29% a year earlier), Codex by 16% and Cursor by 12%. These are self-reported survey figures, and the ranking has moved quickly from one year to the next.

How Agents Choose Tools

The best public evidence on selection comes from vendor-run experiments rather than peer-reviewed work, so it deserves some caution. Armature (September 2026) ran Claude Code, Codex and Cursor across 75 repositories and recorded which tools each agent actually installed. The agents converged on the same tool in only 42% of the cases compared. Some categories have a clear default: Stripe won roughly nine of ten payments picks and Neon took 66% of database picks. Others depend on context. Resend won email in TypeScript repositories while SendGrid won in Python.

Two findings matter more than the league table. First, being mentioned is not being chosen: PayPal appeared 139 times in agent conversations and was never selected. Second, agents research differently. Codex ran a web search in 94% of sessions and Claude Code in about 30%, which means the same vendor is being judged partly on live documentation by one agent and largely on what the model already knows by another. Packaging also had an effect: Supabase lost database picks when it was presented as a bundle rather than as a database.

Build Versus Pick

An agent's cheapest option is frequently to write the feature itself. Amplifying's February 2026 study of 2,430 Claude Code responses found custom code was the most common outcome in 12 of 20 tool categories, including areas such as authentication and feature flags where a human team would often buy. A July 2026 follow-up edition put custom code at 21.4% of picks, and in 32.5% of those cases the agent named a vendor that could be swapped in later. For a vendor, the competitor is as likely to be forty lines of generated code as another company.

Ecosystem and Popularity Bias

Agent choices are not neutral. A University of Zurich study (May 2026) of 13 models found that six of ten provider-affiliated models significantly favoured their own company's ecosystem, by up to 18.8 percentage points in direct questions and 39.2 points in agentic workflows, and that an early ecosystem choice persisted through as much as 90.3% of later decisions. Separately, a peer-reviewed study of library preferences (ACL 2026) found models reaching for older, heavily represented libraries: Flask appeared in 88% of responses against 9% for FastAPI. Both results point the same way. Incumbency in the training data and affiliation with a model provider tilt the field, and a newer tool has to overcome a prior rather than start from zero.

Implications for Developer-Tool Vendors

The practical consequence is that a product's first evaluator is often an agent reading documentation, running an install command and checking whether the result works. That shifts effort toward things an agent can verify: a quickstart that succeeds without a dashboard visit, documentation available as clean text, an accurate llms.txt, a CLI or MCP server for provisioning, and pricing a model can state correctly. This discipline is known as agent experience. None of it guarantees selection, and the evidence base is young, small and largely vendor-produced. What the studies do establish is that agents disagree with each other, that they often build instead of buying, and that familiarity carries weight, so measurement has to be repeated across agents rather than inferred from one.