Are Developers Choosing AI Workflows Instead of AI Models?

The race to build AI coding tools appears to be entering a new phase.

Over the past several weeks, software vendors have announced updates that, taken together, suggest developers are beginning to evaluate AI coding platforms less as isolated assistants and more as integrated development environments built around autonomous software agents.

The shift is evident across several parts of the market.

Last week, Business Insider reported that GitHub CTO Vladimir Fedorov told employees June had been "by far" the company's best month ever, following GitHub's transition of Copilot from a flat-rate subscription model to usage-based pricing. The report noted that the strong results came amid intensifying competition from Cursor, OpenAI Codex, and Anthropic Claude Code.

The pricing change itself may be as significant as the usage growth.

Traditional code-completion tools generated relatively predictable workloads. Today's AI coding agents can perform longer, more computationally intensive tasks that read large codebases, invoke external tools, generate tests, fix defects, and iterate over multiple steps. Charging by consumption rather than by seat reflects that change in how the software is being used.

The same trend is visible across the broader developer tooling landscape.

Microsoft recently made its MAI-Code-1-Flash coding model generally available to GitHub Copilot Business and Enterprise customers, while continuing to expand enterprise policy controls around AI-assisted development. At roughly the same time, Microsoft introduced usage-based pricing for Copilot Cowork, an AI agent designed to execute long-running, multi-step tasks on users' behalf.

JetBrains, meanwhile, released Kotlin 2.4.0 with updates that strengthen its multiplatform development strategy, including support for Java 26, improvements for WebAssembly, and better interoperability across JavaScript and native targets. While not directly an AI announcement, the release continues a broader trend toward development environments designed to span multiple runtimes and workflows rather than a single programming language or platform.

Those announcements follow recent releases from Anthropic, OpenAI, and other AI vendors that increasingly emphasize software agents capable of reasoning over repositories, interacting with developer tools, and completing tasks that extend well beyond code completion.

For software development teams, the distinction may prove important.

The first generation of coding assistants largely focused on generating individual functions, completing lines of code, or answering programming questions. Increasingly, vendors are describing systems that can inspect an application, identify required changes, write code, run tests, revise their work, and continue operating with limited supervision.

The competitive landscape is evolving accordingly.

Rather than asking which model produces the best code completion, development teams are beginning to evaluate broader workflow questions: Which platform understands an existing codebase? Which tools integrate with continuous integration pipelines? Which systems support organizational policies, governance, and security requirements? And which agents can reliably execute multi-step development tasks?

Recent academic research points to the same shift.

In a paper analyzing usage of OpenAI's Codex ("The Shift to Agentic AI: Evidence from Codex"), researchers reported that active users grew more than fivefold during the first half of 2026, more than 10% of users managed three or more concurrent coding agents in a given week, and the share of users assigning Codex tasks estimated to require more than eight hours of human effort increased nearly tenfold. The authors concluded that developers are increasingly using AI to manage longer-running software development workflows rather than simply accelerate individual coding tasks.

Separate research examining more than 180 million public Git repositories (“Detecting AI Coding Agents in Open Source: A Validated Multi-Method Census of 180 Million Repositories”) found growing evidence of AI coding agents participating directly in software development, while suggesting that many current methods underestimate their adoption because agent activity often leaves subtle traces in repositories.

The transition also raises new engineering questions.

As coding agents gain broader access to repositories, shells, and developer environments, researchers are focusing more closely on how those agents are configured and governed. In “A Deterministic Control Plane for LLM Coding Agents,” Padmaraj Madatha argues that the configuration layer for coding agents remains largely unmanaged even as agents receive broad file and shell access, and proposes stronger controls around permissions, audit logs, policy enforcement, and configuration drift.

For enterprise development teams, the result may be a change in how AI tools are evaluated.

The competition increasingly appears to center less on benchmark scores or code completion quality and more on workflow integration. Vendors are racing to build platforms that understand repositories, orchestrate development tasks, work within enterprise governance models, and fit naturally into existing software engineering practices.

If that trend continues, AI coding tools may increasingly be judged not as assistants that help developers write software, but as collaborative development environments that participate directly in the software development lifecycle.

Posted by John K. Waters on June 28, 2026