Hey Reader,

Four posts on Claude Code, all connected. Anthropic shipped Opus 4.7 on Thursday, Claude Code 2.1.112 on Friday, and the thread running through every release this month is the same – make the agent loop tighter.

Here's the reading order I'd suggest.

Claude Opus 4.7 + Claude Code 2.1.112: what actually changed

SWE-bench Verified jumped from 80.8% on Opus 4.6 to 87.6% on 4.7. SWE-bench Pro went from 53.4% to 64.3%. Same price. New `xhigh` effort level that Claude Code now defaults to on every plan. Auto mode available for Max subscribers without a flag.

I recap the benchmarks, the tokenizer change (yes, it's different), and the past two weeks of Claude Code releases in one table.

Claude Code routines – work while you sleep

Anthropic shipped routines the week prior. Saved configurations that run autonomously on their cloud with three trigger types: schedule, API, and GitHub events. If you've been duct-taping GitHub Actions and `claude -p` to get autonomous sessions, routines are the official replacement.

Six practical use cases inside, from nightly backlog triage to cross-repo library porting.

Anthropic's multi-agent code review – what it means for workflow automation

Claude Code 2.1.111 shipped `/ultrareview` (parallel multi-agent code review). That pattern didn't come out of nowhere – Anthropic published the multi-agent architecture back in March, and it's becoming the default for complex agent workflows.

Multiple specialist agents examine PRs from different angles, an aggregator consolidates findings, and the whole thing coordinates through a shared context layer. Worth reading if you want to understand where `/ultrareview`'s design comes from.

WoterClip – stop babysitting your AI coding agent

If multi-agent review is the pattern, WoterClip is the DIY version. An open-source Claude Code plugin that turns Linear into an agent orchestration layer. One Claude instance, multiple personas (CEO, Orchestrator, custom), zero infrastructure.

Persona-based routing lets one Claude handle different types of work based on Linear issue labels. No API server, no database, no separate processes – just markdown files and YAML.

If you want these in order, start with the Opus 4.7 recap (what's new), then routines (cloud automation), then the multi-agent piece (how the pattern works), and finish with WoterClip (running it yourself).

Reply and let me know which one surprised you.

– Alex

P.S. – if you're reading all four of these and thinking "OK I want to actually set Claude Code up this way" – that's exactly what I teach in the workshop.

Production-Grade Claude Code in 5 Days – June 1 cohort. Early bird is $297 through May 15 (saves $200 off the $497 regular price). 30 seats, all still open.

Same setup I use to ship daily, unpacked in 5 days.

Save my seat →