Copilot for JetBrains gets org-curated agents, mid-run CLI steering, and Claude in preview

The first time a teammate slacked me a Copilot CLI invocation and I realised it was halfway through the wrong file, I killed the session and re-ran it from scratch. Twenty minutes of agent work, gone, because the only thing I knew how to do was stop. GitHub's June 22 Copilot changelog post lands a small fix for that exact frustration in JetBrains IDEs, alongside two other moves that, taken together, change how the chat panel feels for anyone using these things inside a regulated org.

A quick map of what shipped, then how each piece lands on a day at work.

What is in the release

Per the GitHub Changelog, custom agents defined at the GitHub organization and enterprise level are now usable directly inside JetBrains IDEs through the Copilot Chat agent picker. Admins publish a curated set, and eligible users get them automatically. Copilot CLI sessions get a mid-flight Send dropdown with three choices: Add to Queue, Steer with Message, and Stop and Send. Claude is available as an agent provider in public preview, plumbed through the Copilot Chat panel. On top of those, the Cloud agent is now generally available, with model picker enhancements and a per-turn AI credits indicator landing in the same release.

Agents the platform team actually owns

The org and enterprise custom agents are the change I expect to hit most teams the fastest. Until now, the question of whether a platform team could standardise on a set of agents and stop every developer from rolling their own had a fairly ugly answer. You could write a wiki page and hope. With this release, admins publish the agents and JetBrains users see that curated list in the agent picker, and the changelog describes automatic distribution to eligible users, with no per-machine install dance.

If you have spent any time on a platform team in 2026, you know why this matters. The agent your developers use is now part of your build supply chain. Treating it as shared infrastructure is overdue.

A small dropdown that changes how you script a long run

The CLI change is the one I am most quietly happy about. While a Copilot CLI request is running, the Send button turns into a dropdown. Add to Queue lines up your next message behind whatever the agent is currently doing. Steer with Message yields after the active tool execution finishes, then processes the new message. Stop and Send abandons what was running and starts fresh.

In day-to-day terms: you can hand the agent a long task, watch the first tool call land, and as soon as you spot the wrong directory, send a correction without killing the run and losing context. For non-interactive scripts wrapped around the CLI, this changes the shape of what is reasonable to write, because partial state is now recoverable.

Claude as a provider, in preview

This is the door-opener. Per the post, you install the Claude Code CLI locally, point Copilot at the binary under Settings, Tools, GitHub Copilot, Chat, and Claude then appears in the agent picker inside Copilot Chat. It is public preview, so I would not yet wire compliance reporting against it, but the shape of the thing is what people have been asking for: the Copilot surface, with the agent engine underneath chosen per session.

One quieter piece worth flagging alongside this: the Agent Debug panel has a new logs summary view with aggregate stats for the session, reached by selecting the session name from the session list. If you are tracking what your agents actually did across a long debug, that view is the missing breadcrumb.

The rough edges

I am not pretending all of this is finished. Custom agents at org level are a governance primitive without a clear answer yet for how you version them or roll one back when one starts misbehaving. The CLI mid-flight steering is great in an interactive session, less obvious in a CI script where there is no one to hit the dropdown. And Claude as a provider being in preview means you should expect breaking changes to the install path before it goes GA.

What I am watching next

Two things. First, whether the curated org-agents picker grows an audit trail, so a security review can ask "who published this agent, when, with which prompt", and get a real answer instead of a screenshot. Second, whether the Steer-with-Message verb finds its way into automation: a non-interactive way for a parent process to steer a child agent run, the same way humans now can, would quietly change how long-running agentic jobs get composed in pipelines.