Junie hits GA: what changes for a developer who lives in the agent loop

What it is

Junie is JetBrains' autonomous coding agent. It takes a task, plans the steps, edits across files, runs your tests, and iterates until things pass — either inside a JetBrains IDE or from a standalone CLI. As of June 17, 2026, JetBrains moved Junie out of beta, and the headline GA feature is that Junie can drive your IDE's debugger the way you would — starting or joining a debug session, launching a run configuration, debugging a test, or taking over a session you already have open.

The Monday-morning test

You open IntelliJ on Monday to a flaky integration test that failed overnight. The traditional agent loop is: agent adds println, reruns, guesses again. Junie's pitch is different — instead of sprinkling println and guessing, it can set breakpoints, step through execution, inspect variables, and iterate on fixes against live runtime state. If that holds up on your codebase, you delegate the test, go make coffee, and come back to either a green run or a real stack trace explanation — not a pile of speculative log lines to clean up.

That's the use case worth testing first. Not greenfield feature generation — debugging a known-failing test.

Pricing

Junie ships under the JetBrains AI subscription. According to JetBrains' pricing page, each AI Credit corresponds to $1 USD charged in your local currency, and the number of credits included in your plan corresponds to its subscription price, with a small bonus for AI Ultimate. The free tier is thin: with the AI Free tier you get 3 AI Credits every 30 days, and that is your full monthly cloud quota for JetBrains AI (AI Assistant and Junie).

Two important escape hatches confirmed in JetBrains' AI FAQ: you can bring your own API keys from OpenAI, Anthropic, and other OpenAI-compatible providers in AI Chat, Junie, and Claude agents, and no JetBrains AI subscription is required to use BYOK for chat and agents. So you can run Junie on your own model budget. Cloud-completion features still need a paid plan.

Third-party reviews quote paid tiers starting around $100/user/year, but those are not first-party — treat them as claims, not facts, and check the JetBrains AI plans page for your seat count.

What we'd actually use it for

Narrower than the vendor pitch:

• Debugging a single failing test with the IDE debugger in the loop — the one task where Junie's GA story is genuinely differentiated.
• Mechanical refactors with verification — rename across modules, extract a service, then let Junie rerun the suite.
• CLI runs in CI for codemod-style changes — Junie also runs from the terminal and from GitHub issues and pull requests, and from GitLab issues or MRs, which makes it usable as a job rather than an interactive partner.

What we would not lean on it for yet: greenfield architecture decisions, anything touching SDK surface area where you need a human to own the contract, or eval harness design.

Limits

• Credit burn is real. The community thread linked above documents serious user concerns about credit consumption on long agent runs. Multi-file refactors in a paid agent loop eat credits faster than chat does. Budget accordingly or use BYOK.
• JetBrains IDEs only for the IDE experience. The agent's debugger integration assumes you're already in IntelliJ, PyCharm, WebStorm, GoLand, PhpStorm, RubyMine, RustRover or Rider. VS Code shops get the CLI, not the debugger story.
• Still an agent. Plans go sideways on large unfamiliar repos. The test-and-fix loop only works if your tests are fast and meaningful — if your suite takes 20 minutes, the loop is unusable.
• Model choice matters. Per the Junie GitHub repo, Junie is an LLM-agnostic coding agent — which means quality varies meaningfully with which model you point it at, and that's on you to evaluate.

Try it if

• You already pay for a JetBrains IDE and have a real failing-test backlog.
• You want to keep agent runs inside the IDE rather than juggling a separate tool window.
• You have BYOK access to Anthropic or OpenAI and want to control spend.
• You need the same agent reachable from CI via a CLI.

Skip it if

• You live in VS Code or Neovim and don't want a second IDE just for the agent.
• Your test suite is too slow for a test-and-fix loop to converge.
• You're on the free tier and expect to run real agent tasks — 3 credits a month won't cover it.
• You need deterministic, auditable refactors. An agent loop is the wrong tool; write the codemod.