JetBrains names the debt AI agents leave behind

Developers have spent years fighting technical debt. Now there’s a new variant accumulating in codebases, and many teams don’t even know about the pileup.

JetBrains calls it Shadow Tech Debt — low-quality, architecture-blind code generated by AI agents that operate without any structural understanding of the projects they’re modifying. And with the launch of Junie CLI on Monday, the company is betting that this is the next big problem in enterprise software development.

The diagnosis starts with fragmentation.

“The current state of working with coding agents is fragmented: Each agent runs in a separate tool, with a different setup, different context, and no structural understanding of your code,” writes Nik Tkachev, a head of product at JetBrains, who announced JetBrains Air this week in a blog post. JetBrains argues that it is code that works in isolation but quietly undermines the coherence of the broader codebase over time.

As agentic coding tools have proliferated over the past 18 months, engineering teams have embraced them for the productivity gains — and largely deferred the question of what happens to code quality at scale. Agents don’t read architecture decision records. They don’t know why a particular pattern was chosen three years ago. They don’t understand the tribal knowledge baked into a legacy module. They complete the task at hand and move on.

That’s a reasonable tradeoff for a one-off script. It’s a more serious problem when autonomous agents run in CI/CD pipelines, spin up pull requests, and commit directly to shared repositories.

“Let’s be honest: Complex codebases aren’t yet ready for pure agentic coding,” Tkachev writes. “Air focuses on agent orchestration without replacing existing development workflows.”

Enter Junie CLI

Junie CLI, now in beta, is JetBrains’ more direct answer to the code quality problem. The tool is designed as a fully standalone coding agent — available in the terminal, in any IDE, on GitHub, on GitLab, and in CI/CD pipelines — with what the company describes as “codebase intelligence” at its core.

“Junie isn’t just ‘AI in a terminal,'” writes Anastasia Krivosheeva, head of marketing excellence at JetBrains, who announced Junie CLI in a separate blog post. “It’s a fully standalone agent with capabilities designed to move beyond simple prompting.”

Rather than treating each task as a stateless prompt, Junie CLI incorporates structured project context and workflow awareness, so the code it generates is grounded in the actual shape of the codebase.

The product also includes next-task prediction — the agent anticipates what a developer might need based on project context — and one-click migration from competing tools, including Claude Code and Codex. It supports models from OpenAI, Anthropic, Google, and Grok, and uses a BYOK (bring your own key) approach with no additional platform charges beyond model costs.

Agentic coding at scale

“We hold a principled optimism for agentic software development — and a pragmatic one,” Tkachev writes. “New concepts are emerging faster than anyone can validate them, so we’d rather ship what works than hype what might.”

JetBrains is telling developers that agentic coding at scale requires professional-grade infrastructure, without alienating developers already satisfied using Claude Code or Cursor. The Shadow Tech Debt framing does that work — it doesn’t attack the agents themselves, it attacks the conditions under which they operate.

JetBrains says the agentic era needs a neutral infrastructure layer rather than yet another agent.

“We’re expanding from IDE-native AI to ecosystem-level AI — using one agent to connect platforms,” Krivosheeva writes.

Junie CLI launches alongside JetBrains Air, a new agentic development environment built on the open Agent Client Protocol, which lets Claude Agent, Gemini CLI, Codex, and Junie run side by side in a single workspace. Together, the two products position JetBrains less as a competitor to any individual agent and more as the platform layer beneath them all.

“Junie meets you where you are. By making Junie available outside JetBrains IDEs, we’re expanding from IDE-native AI to ecosystem-level AI – using one agent to connect platforms,” Krivosheeva writes. “This is a significant milestone for us and an important step toward enabling professional-level development, even outside the IDE.”