GitHub added new Copilot features for JetBrains IDEs, including Claude as an agent provider preview. Here is what developers should evaluate first.
What actually shipped
GitHub has expanded Copilot inside JetBrains IDEs with new agent-oriented features, including a preview that lets Claude act as an agent provider. That is more than another chat model dropdown. An agent provider is meant to plan work, call tools, and drive multi-step edits inside the IDE instead of only answering a single prompt. For JetBrains users who already live in IntelliJ, PyCharm, or WebStorm, the practical shift is that agent-style workflows can stay closer to the project you already have open—files, run configurations, refactors, and tests—without bouncing to a separate browser session.
Because this is a preview, treat it as an evaluation surface rather than a default for production-critical work. Preview features often change behavior, tool access, and model routing without much notice. Your job is to learn where the agent is reliable, where it needs tight supervision, and which tasks still belong to you.
What to evaluate first
Start with tasks that are bounded, reversible, and easy to verify. Good first candidates: explain a confusing module, propose a small refactor with a clear acceptance check, draft unit tests for an existing function, or walk through a failing stack trace and suggest a minimal fix. Avoid handing it large, ambiguous goals like “modernize this service” until you have seen how it reads your codebase, preserves project conventions, and stops when uncertain.
- Context quality: Does it use the right files, packages, and project structure, or does it invent paths and APIs?
- Edit discipline: Does it make focused diffs, or rewrite whole files and ignore local style?
- Tool use: When it runs searches, builds, or tests, does it interpret failures usefully and iterate?
- Handoff cost: How much cleanup do you still do after accepting suggestions?
Score each session on those points instead of on how polished the first answer looks. An agent that asks one clarifying question and ships a small correct patch is more useful than one that produces a long plan and a brittle multi-file change.
Workflow habits that keep you in control
Use a tight loop: state the goal, constraints, and definition of done; let the agent propose a plan; approve the plan before edits; review the diff file by file; run your own tests. Keep secrets, credentials, and production configs out of the agent’s working set. Prefer feature branches and small commits so you can discard a bad trajectory quickly. If your team already has coding standards, lint rules, or architecture docs, point the agent at those explicitly—previews rarely infer team norms on their own.
Also decide when Claude-as-agent is the right tool versus plain completion or a short chat. Completions still win for local typing and boilerplate. Short chat still wins for “what does this API mean?” Agent mode earns its keep when the task spans several files, needs a sequence of checks, or benefits from iterative repair after a failed test run.
Team and risk checks before wider adoption
Before rolling a preview agent provider across a team, agree on review rules. Every agent-authored change should get the same code review as a human PR: correctness, security, performance, and maintainability. Watch for over-trust—agents can produce confident, well-formatted code that still misses edge cases or license constraints. Track a few simple signals over a week of real work: time to a mergeable patch, number of review comments, and how often you revert agent edits. If those numbers do not improve on the tasks you care about, keep the feature available for experiments but do not make it the default path.
The useful takeaway is not that JetBrains Copilot now “has agents.” It is that you can test Claude as an agent provider on real project work, with your own toolchain and quality bar. Evaluate on scoped tasks, keep humans on the approval path, and expand usage only where the agent consistently reduces rework rather than moving it into review.