Does a background review differ in quality from one that used to run directly in the conversation?
A background Subagent is still, fundamentally, an independent Claude instance with its own system prompt, tool list, and model configuration — the review logic itself hasn't changed just because of running in the background. The difference is mainly in how the process is presented: during background execution, the intermediate steps of reading and analyzing each file aren't shown live in the main conversation, and only the final review conclusion gets returned.
What actually matters here is an information gap, not a quality gap. If you're used to judging how thorough a review is by watching its intermediate steps — the order it reads files in, its hesitation over a particular piece of code — those details aren't directly visible in background mode; you only see the final result, which is a real change in experience for anyone who wants to trace the review's reasoning process.
Why did Anthropic make /code-review background-only now, and does the timing relate to other updates?
This change isn't an isolated event — it's one piece of a broader, ongoing round of improvements to Claude Code's Subagent system. The same batch of updates also raised the default nesting depth limit for subagent spawning, removed the 200-subagent-per-session cap, and patched several permission and isolation-related bugs affecting background subagents (for example, worktree-isolated sessions that could previously run destructive git commands against the main checkout). Together, these changes point in one direction: Claude Code is treating background parallel execution as the default mode for handling time-consuming tasks, not a special-case feature.
Seen in that context, moving /code-review to background execution reads more like a natural extension of this broader direction than a standalone fix aimed only at the "conversation gets flooded" experience problem. The practical implication of understanding this is that other time-consuming commands (more complex analysis, multi-step workflows) will likely gradually migrate to the same background execution model going forward — a product direction worth watching, not just this single update.
If I need to watch each step of the review process live — for teaching or debugging purposes — is that still possible?
Based on what's publicly known, /code-review now defaults to the background Subagent path, meaning the default behavior is background execution with no live display of the process. If a situation genuinely calls for watching the step-by-step process live — say, demonstrating to a teammate how Claude reasons about a code issue, or debugging whether the review logic itself is running correctly — a more practical alternative is manually triggering review logic through a foreground approach that blocks the main conversation and shows the process step by step, rather than relying on the default /code-review command.
For most scenarios that don't need live process tracking and only care about the final review conclusion (day-to-day PR review, for instance), the experience improvement from background execution is a direct win. But if your workflow has relied on live visibility into the review process, this is an area where you'll need to adjust your habits after this update.
Will this change affect review-related automation I've already set up, like a CI script that calls code review?
For automated workflows that invoke /code-review via the CLI or a script, the main thing to verify is whether that workflow assumes the review process streams output live to a log or screen (for instance, relying on step-by-step output to determine whether the process is stuck). If the automation logic waits for the final result rather than parsing intermediate step output, this change shouldn't in principle break the existing workflow — the timing of the return simply concentrates at the moment the review completes, rather than streaming step by step.
The safer move is to actually run your existing automated workflow once and confirm the format and timing of the returned result match expectations. In particular, if your workflow has a timeout mechanism configured, background execution itself typically doesn't lengthen the total execution time — but if the original timeout logic was tied to "how long since the last new output" rather than "total execution time," that kind of logic may need re-examining under background mode.
Anthropic's August 2026 Claude Code update changed how the /code-review command executes: the command now runs as a background Subagent, with the review process getting its own independent context window instead of consuming the main conversation step by step — the review result gets returned once it finishes.
This addresses a concrete problem many developers have actually run into: previously, running /code-review meant every file Claude read and every step it analyzed showed up in the conversation in full, and reviewing a medium-to-large PR would often flood the current conversation with a mass of file-reading logs. That not only made the screen harder to follow, but also meant a significant chunk of that conversation's context got consumed by the review process itself, squeezing out room for continuing the discussion afterward.
According to the official changelog, this change also targets "stacked slash commands" as the review target — meaning the underlying review logic itself was adjusted alongside this change. More broadly, Claude Code's current subagent system supports two modes: foreground (blocks the main conversation until it finishes) and background (runs concurrently alongside the main conversation). Before a background subagent launches, it asks what tool permissions it will need; once running, it operates within that approved scope, and any operation that wasn't pre-approved is automatically denied. This update makes /code-review default to the background path, letting users keep working on other tasks in the main conversation without having to wait for the review to finish first.
Worth noting: local /code-review runs follow a project's CLAUDE.md configuration but do not read REVIEW.md — an easy configuration mismatch to overlook if a team relies on both managed review and local review and expects both to enforce identical rules.
If you're used to asking Claude to review code in the same conversation where you're writing it, this change directly reduces the context-consumption problem — previously, reviewing a larger PR often meant needing to start a fresh conversation just to keep discussing next steps cleanly, and background execution reduces how often that switch is necessary. One practical thing to watch: a background subagent's result only comes back once it's done, so if you keep asking other questions on the main thread while waiting, remember to check back on whether the review has finished so you don't miss an important finding. Also, if your team relies on both local and managed review paths, it's worth taking time to confirm that CLAUDE.md and any team rule files are actually configured consistently across both, so the standards actually being enforced don't quietly drift apart.