Every Claude & Claude Code feature shipped in March 2026: Code Review agents, Auto Mode, 1M context GA, Computer Use on Mac, Mythos leak & Code with Claude c...
Agentic Code Review and Auto Mode
Two of the month's releases pushed Claude Code further toward running work on its own. Code Review agents read a diff the way a human reviewer would: they trace what changed, flag correctness risks and reuse opportunities, and report findings you can act on before merging. Instead of pasting a patch into a chat window, you point the agent at a branch or a pull request and let it do the reading.
Auto Mode complements that by letting Claude carry a multi-step task through to completion with less turn-by-turn prompting. The practical tradeoff is oversight: the more autonomy you grant, the more you need clear success criteria and guardrails around anything that writes to shared systems. Treat Auto Mode as a way to compress the mechanical parts of a task, not a reason to stop defining what "done" means.
1M Context Reaches General Availability
The 1M-token context window moved to general availability, which matters most for work that used to require chopping inputs into pieces. Whole repositories, long design documents, and extended transcripts can now sit in a single request, so the model reasons over them together instead of losing the thread between chunks.
A larger window is not a license to dump everything in. Longer prompts cost more and can bury the signal you actually care about. Some ways to use the headroom well:
- Load an entire codebase when a change touches many files and cross-references matter.
- Keep long-running agent sessions coherent without aggressive summarization.
- Analyze large logs or documents in one pass rather than stitching partial answers together.
- Still trim obvious noise — relevance beats raw volume, even with room to spare.
Computer Use Comes to Mac
Computer Use on Mac lets Claude operate a desktop directly — reading the screen, moving the pointer, and clicking through interfaces the way a person does. That opens up tasks with no clean API: driving legacy apps, filling forms across tools, or automating a workflow that only exists as a sequence of clicks.
Because the model is acting on a real machine, the safety posture changes. Run it against accounts and environments where a mistaken action is recoverable, watch what it does on consequential steps, and scope its access rather than handing it your primary desktop. The capability is powerful precisely because it can touch anything on screen, which is also why it deserves boundaries.
Mythos, Code with Claude, and What to Watch
The month also brought the Mythos leak and the Code with Claude event into view, both hinting at where the tooling is headed. Leaks are worth reading with skepticism — details shift before anything ships — but they signal the direction of investment, and a developer-focused gathering is a clearer read on the roadmap teams are being asked to build against.
Taken together, March's releases point the same way: less time spent shepherding the model through individual steps, and more spent setting up the task, the context, and the checks around it. The useful skill now is deciding what to delegate, giving the agent enough context to succeed, and keeping verification tight enough that autonomy stays an asset instead of a liability.