Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out

TEXT START: As agents take over more and more parts of the coding journey, there is one specific part everyone outsources to their agent, from vibe coders with no software background to senior engineers: selecting which service to implement for a specific need in an existing codebase.

The Dissection

The article is documenting the transfer of software procurement from humans to coding agents. Tool selection becomes an agent-mediated distribution market: vendors are no longer competing primarily for developer attention, but for inclusion in the model’s recommendations and implementation patterns.

The 17,000-run experiment is presented as empirical validation that agents repeatedly converge on particular tools across personas, repositories, prompts, and coding systems. Its real subject is not databases or developer convenience. It is the emerging power of agents to determine which firms receive revenue and which become invisible.

Under the Discontinuity Thesis, this is early evidence for P1: cognitive work is being automated. It also exposes the beginning of P2: human institutions are losing control over a formerly human-controlled economic gate. It does not, by itself, prove P3. It shows the mechanism forming, not the entire labor market already collapsing.

The Core Fallacy

The article conflates agentic recommendation with reliable judgment and market legitimacy.

An agent choosing Neon repeatedly does not prove Neon is objectively the best database. It proves that, under the tested prompts, repositories, model priors, discovery channels, integration patterns, pricing assumptions, and evaluation criteria, Neon was the easiest answer for the agents to produce and implement.

The experiment measures behavioral convergence, not truth. Agents may optimize for familiarity, training-data frequency, installation simplicity, API ergonomics, or learned vendor narratives. A tool can win the recommendation layer while losing on long-term reliability, security, compliance, lock-in, operational cost, or resilience.

The deeper fallacy is more consequential: treating this as a new productivity advantage for developers rather than evidence that developers are being demoted from decision-makers to approval interfaces. The human still clicks “approve,” but the economically valuable cognition has already moved upstream into the agent.

Hidden Assumptions

• The tested prompts accurately represent real production requirements.
• Three agents, 75 repositories, and 1,163 prompt variations generalize to the wider coding-agent market.
• Repeated recommendations indicate product quality rather than model bias or discoverability advantage.
• The agent receives enough context to evaluate security, compliance, migration risk, total cost, and operational durability.
• Successful implementation means successful technology selection.
• Vendors cannot manipulate agent rankings through documentation, integrations, pricing, benchmark gaming, or model-targeted marketing.
• The current agent behavior will remain stable as models, tool catalogs, and commercial incentives change.
• Human approval remains meaningful rather than becoming ceremonial.
• A vendor that wins agent recommendations will retain customers after the initial implementation and survive the resulting concentration of power.
• The coding-agent transition creates more valuable human roles than it destroys.

The sample size is impressive as an observation of current behavior. It is not a warrant for universal validity. Volume does not cure selection bias; it merely gives the bias a larger scoreboard.

Social Function

Classification: partial truth, transition management, and prestige signaling, with an ideological-anesthetic effect.

The partial truth is real: coding agents are becoming procurement agents, implementation agents, and distribution channels. Vendors that ignore this will be erased from consideration before a human ever evaluates them.

The anesthetic is the framing. Structural displacement is repackaged as a leaderboard and an optimization problem. The article invites developers and vendors to ask which tool the machine prefers, rather than asking what happens when the machine no longer needs most of the people who used to make those decisions.

Its most valuable audience is not the individual developer. It is the vendor class, which needs to understand how to become machine-legible, machine-discoverable, and machine-preferred. The human programmer appears as the customer, but increasingly functions as the final authorization layer on an automated commercial pipeline.

The Verdict

This is not evidence that developers remain sovereign. It is evidence that sovereignty is migrating into the agent layer.

The experiment captures a live transition in which coding agents begin deciding who gets installed, paid, and remembered. That is a genuine early-stage confirmation of the Discontinuity Thesis. But the article mistakes the efficiency of the transition for its social outcome. Better tool selection does not preserve productive human participation; it accelerates its removal.

The likely winners are the Sovereigns controlling models, distribution, compute, proprietary data, and agent defaults. A narrow Servitor class may remain valuable in security, architecture, governance, and high-consequence exception handling. Everyone else risks becoming a human rubber stamp attached to a software procurement machine.

The headline is therefore accurate but incomplete: the agents are choosing the tools. The next question is how long they continue needing the people who currently believe they are the ones choosing.