The AI Agent Race Is On. But Are We Watching the Right Race?

In the rapidly evolving landscape of software development, AI coding agents have surged into the mainstream. A new Techstrong special report, "The AI Agent Race: At the Top of the Stretch," examines the competitive field of AI agents, the metrics used to measure their success, and what enterprises should consider before making their choices.

While multiple players are vying for dominance, there is no single universal market-share table for AI coding agents, as the numbers measured vary significantly. Claude Code, OpenAI Codex, GitHub Copilot, Cursor, and other challengers are competing for the future of software development. To better understand the situation, the report begins by separating the different races within the AI coding agent market.

The report highlights the immense popularity of AI coding agents among professional developers, with 90% using them at work at least weekly and 68% using them daily. Claude Code emerged as a notable player, with 39% of respondents using it at work, up from 18% in January. However, other evidence suggests a different picture, with Anthropic's share at the foundation-model level being considerably lower than Claude Code's position in developer surveys.

Enterprises use AI coding agents in various ways, depending on their specific needs. For instance, they might use OpenAI models in customer-facing applications, Google Gemini for data analysis, and Claude Code for software development. Additionally, an enterprise might access Claude through GitHub Copilot or Cursor rather than through Anthropic's own interface.

The report also points out that the line between model, agent, development environment, and enterprise platform is becoming increasingly blurred, as they are not always purchased from the same vendor. This complexity adds to the challenges of evaluating and comparing AI coding agents. One of the key findings is that most enterprise adoption remains concentrated in assisted rather than autonomous work.

Assistance accounts for 47.20% of AI use, while semi-autonomous agents represent 13.59%. Autonomous agents, which aim for end-to-end automation, represent only 5.84% of AI adoption. Enterprises are willing to let AI help create and review changes but become more cautious as agents approach infrastructure, credentials, and production.

This cautious attitude is necessary due to the potential consequences of failure when dealing with secrets, infrastructure, or production deployment pipelines. The report examines the often-confusing landscape of benchmark scores, which can create a false sense of precision. Benchmark scores depend on multiple factors, such as the model, agent harness, available tools, task design, execution environment, and resources allowed for consumption.

Even minor changes, like using a familiar public repository instead of an internal codebase, can significantly impact benchmark results and give enterprises little insight into how the agent will perform on their own work. The report also discusses the cost question surrounding AI coding agents, which can be as confusing as performance metrics.

Pricing models vary, including monthly seat prices, token rates, and premium-request allowances. All these factors make it challenging for enterprises to make informed decisions when selecting the right AI coding agent for their needs.