Anthropic used its second annual Code w/ Claude developer conference on May 6 to push Claude Code well past its “AI pair programmer in a terminal” origins. The headline was the general availability of Claude Opus 4.7 — the model that powers this very assistant — but the surrounding announcements are the bigger story: the company is repositioning Claude Code as a background agent platform you delegate to, not a chat session you babysit.
What actually shipped:
- Claude Opus 4.7, GA. Anthropic’s flagship model is now generally available with improvements specifically targeted at long-horizon software-engineering work and stronger vision capabilities. A 1M-context tier of Opus 4.7 is what Claude Code itself runs on for large-codebase sessions.
- Remote Agents. You can now hand Claude Code a task from your phone and have it execute on a cloud machine you’ve connected — your laptop is no longer the substrate. The pitch: kick off a refactor on the commute, review the diff at the desk.
- Code Review with CI auto-fix. Claude Code can now review pull requests and push commits that resolve its own findings. The CI integration closes the loop most existing AI review tools leave open: they comment, then a human applies the fix.
- Routines. Recurring workflows can now be scheduled or triggered by webhooks — security audits, dependency bumps, broken-link sweeps — without a developer in the loop.
- Multi-agent orchestration. Specialized sub-agents (planner, implementer, reviewer) can be composed for a task, with Claude Code acting as the orchestrator. This is the same architecture Anthropic uses internally; they’re now exposing it as a first-class primitive.
- Redesigned desktop app. Drag-and-drop workspaces, a session sidebar, an integrated terminal and file editor, faster diffs, and SSH support on Mac. Parallel sessions are the design center — running four agents at once is no longer a power-user pattern.
Two structural changes behind the announcements are worth pulling out:
- Usage limits roughly doubled across Pro, Max, Team, and seat-based Enterprise tiers, and peak-hour throttling on Pro and Max is gone. Anthropic separately disclosed a compute partnership with SpaceX that gives them substantial new capacity — the rate-limit move is what that compute pays for in the near term.
- The “agent that runs while you sleep” framing is now official. Routines + Remote Agents + multi-agent orchestration are not three independent features. They are the same product thesis stated three ways: developers shouldn’t be the bottleneck for tasks the model can finish autonomously. That thesis collides directly with Codex (OpenAI) and Gemini Code Assist, both of which are still anchored to interactive sessions.
What this likely means downstream:
- CI vendors get squeezed. Auto-fix is a feature CircleCI, GitHub Actions third-party marketplaces, and standalone code-review SaaS have been racing to ship. Anthropic shipping it natively into Claude Code — at no per-fix charge — changes the buy-vs-build math for engineering orgs already paying for seats.
- Scheduled agents are the new cron. Routines is, in form, a cron that can think. Expect a wave of internal-tooling teams to migrate “weekly maintenance” scripts to Claude Code routines and absorb the LLM cost as a line item against the headcount it doesn’t have to hire.
- The model lock-in question gets sharper. Opus 4.7’s edge on long-horizon coding tasks is reportedly meaningful, but the surrounding system — routines, IDE plugins, Remote Agents, the desktop app — is what makes leaving expensive. The fight in 2026 is increasingly platform, not weights.
The catch, where there is one:
- Remote Agents and routines depend on shared infrastructure that historically has had degradation incidents — and the partial outages the platform saw through Q1 are exactly the kind of thing that scares teams away from putting cron-grade workloads on it.
- The “AI fixes its own PR comments” loop is a clear productivity win in greenfield code and a clear footgun in any codebase with non-obvious invariants. The early adopters who get this right will have spent serious effort on agent permissions and CI gating.
- Multi-agent orchestration is still expensive in tokens. Pricing realism on long multi-agent runs will be the real adoption gate, not the demo.
The keynote replay and full session recordings are live on Anthropic’s site; the changelog has the per-feature flags.
Sources