Claude Code March 2026 Updates Add Code Review and Auto Mode

A month of updates aimed at the slow parts of coding

In March 2026, Anthropic put out a long run of Claude Code releases. Taken together, they say a lot about where AI coding assistants are going. The individual features matter, but the pattern matters more. Most of the month's work went into the parts of software development that come before and after writing code: approving actions, scheduling recurring jobs, watching sessions remotely, and above all reviewing pull requests. GitHub spent the same month rebuilding Copilot around the same problem. That makes March the clearest sign so far that the competition among coding assistants has moved from producing code to checking it.

One count has the first half of the month running from version 2.1.63 through 2.1.7613. An archived changelog lists releases continuing to v2.1.88 by March 3120. The pace alone shows how hard Anthropic is pushing the product.

Code Review: Anthropic goes after the review backlog

The launch with the biggest long-term consequences was Code Review. Anthropic introduced it on March 9 as a research preview for Team and Enterprise plans21. When a pull request opens, the system sends several agents to examine the diff and the code around it in parallel. A verification pass then filters out false positives, and the remaining findings are ranked by severity before they are posted2326. Anthropic justified the feature with its own numbers. It said code output per engineer had grown 200% in a year, and that the share of internal pull requests getting substantive review comments rose from 16% to 54% after deployment21. The company also said engineers marked fewer than 1% of findings as incorrect, and that an average review takes about 20 minutes23.

The coverage agrees on the overall design but not on the details. One breakdown describes exactly five specialized agents29. InfoQ reports that the number of agents grows with the size and complexity of a pull request23. Anthropic's open-source /code-review plugin, a related but separate tool, launches four parallel agents and only reports issues with a confidence score of 80 or higher22. The most likely explanation is that the hosted product adjusts the number of agents dynamically, and fixed counts come from simplified descriptions or from the plugin.

Cost is the main objection. Reviews are billed by token usage and average roughly $15 to $25 per pull request30. For a 15-person team opening 100 pull requests a month, one analysis puts the total at $1,500 to $2,50027. InfoQ reported that some developers doubted that price and the 20-minute turnaround would work for high-volume teams23. Anthropic has kept one design choice firm: Code Review never approves a pull request, so the merge decision stays with a person23. The feature became generally available to Team and Enterprise organizations on October 15, along with promotional review credits for eligible Enterprise customers21. That means the March preview held up through seven months of real use.

GitHub Copilot shipped a competing reviewer four days earlier

GitHub got there first. On March 5, Copilot code review moved to an agentic, tool-calling architecture and became generally available to Copilot Pro, Pro+, Business, and Enterprise users1. Instead of reading only the diff, the reviewer now pulls in related code, directory structure, and references so its comments account for the wider codebase1. GitHub says usage has grown tenfold since the feature first launched, and that Copilot now handles more than one in five code reviews on the platform6. In 71% of reviews the agent left actionable feedback, averaging 5.1 comments per review where it commented at all5.

The two companies made very different choices. Copilot's reviewer runs on GitHub Actions, so organizations that have opted out of GitHub-hosted runners have to set up self-hosted runners once1. Anthropic's runs on its own infrastructure with read-only repository access, and does not run in self-hosted environments21. Copilot's comments never count toward required approvals5, which matches Anthropic's no-approval rule. GitHub is selling convenience and volume inside the platform where developers already work. Anthropic is selling depth and charging per review for it.

Third-party comparisons need caution. One guide rates Copilot at about 55% bug detection using a "single model pass"29, which doesn't match GitHub's own description of an agentic, multi-step architecture1. Accounts of which plans include review also differ. GitHub's changelog includes Pro and Pro+1, while one independent guide says only Business and Enterprise qualify3. GitHub's own announcement is the more reliable source here.

Copilot's other March changes point the same way. The coding agent now reviews its own changes with Copilot code review before opening a pull request7. It also starts work 50% faster10. GitHub's monthly digest added that developers can request reviews from the command line, and acknowledged a rate-limit incident that disrupted many users' workflows during the month8.

Claude Code works longer without supervision

Outside of review, most of Anthropic's March releases make Claude Code run for longer stretches without a person watching. The /loop command arrived on March 7 with cron scheduling, so prompts can repeat on an interval during a session13. Scheduled tasks can also run recurring jobs on Anthropic-managed cloud infrastructure11. On March 20, a --bare flag for scripted calls shipped alongside a --channels permission relay that can send tool-approval prompts to a phone14.

Auto mode is the most significant of these. It entered research preview for Team plans on March 24, using a classifier to let routine actions through and block or escalate risky ones19. Anthropic's documentation describes it as a middle ground between approving every action and turning off permission checks entirely12. It became generally available on July 1019. Computer use came to Claude Code and Cowork for Pro and Max users on March 23, letting Claude open files, run developer tools, and click through the screen11. Cloud auto-fix lets web and mobile sessions follow pull requests, fix failing CI runs, and respond to review comments11. That links the coding agent directly to the review loop.

Changes to the model, context, and enterprise controls

The changes to the model underneath were quieter but had real effects. On March 4, Opus 4.6 began defaulting to medium effort for Max and Team subscribers. The "ultrathink" keyword returned for high effort, and Opus 4 and 4.1 were removed, with pinned users moved to 4.6 automatically14. On March 13, the 1M-token context window became the default for Opus 4.6 on Max, Team, and Enterprise plans, where it had previously required extra usage14. Reports differ on timing: one writer says the 1M context launched in February15. The changelog's wording suggests both are true. The capability existed earlier, but it became the default in March. Output limits went up as well, to a 64k default and a 128k ceiling for Opus 4.6 and Sonnet 4.611.

The medium-effort default was controversial. A widely shared DEV Community post argued that adaptive thinking plus the lower default effort hurt performance on complex engineering work. It described Anthropic offering an environment variable as a temporary workaround while it investigated17. It is hard to judge how widespread the problem was from one community account. Still, it shows the tension in making medium effort the default: cheaper, faster turns can come at the cost of correctness on difficult problems.

Enterprise administrators got several tools. A managed-settings.d/ directory lets separate teams deploy their own policy fragments. A sandbox.failIfUnavailable setting makes Claude Code stop with an error when the sandbox can't start, instead of quietly running without it. A PowerShell tool arrived for Windows as an opt-in preview14. Early-March releases also fixed a bug that let .mcp.json servers skip their trust dialogs on first run, and cut prompt re-renders by 74%15.

What it adds up to

The main takeaway is that Anthropic and GitHub reached the same conclusion in the same week. Generating code has gotten cheap, and checking it is now the bottleneck. GitHub is building review into the subscription and running it at platform scale. Anthropic is charging per review and betting that a deeper, slower, multi-agent pass is worth paying for. Neither lets the AI approve a merge. One older comparison chart says Claude Code has no native review4, and it was outdated within days of publication.

For teams evaluating tools, the choice now looks less like picking one assistant and more like combining layers: one agent writes the code, another reviews it, and a person decides what ships. March is when both companies started selling that arrangement.