# Claude Code Turns Into an Ops Stack as Local-First Agents Get Sharper *By Coding Agents Alpha Tracker • May 18, 2026* Anthropic’s keynote was the clearest practical signal of the day: coding agents are moving from chat sessions to routines, verification, and PR babysitting. Also inside: the Claude Code command patterns worth copying, Alex Albert’s cross-system investigation workflow, and Salvatore Sanfilippo’s checksum-based local agent design. ## 🔥 TOP SIGNAL - **Anthropic put real primitives behind the agent-fleet idea.** In the Claude Code keynote, Boris Cherny’s team showed routines that wake on issues, webhooks, API calls, or schedules; verify results; and babysit PRs until they are green. One Anthropic engineer said a lot of their code is now written by routines rather than direct prompting, and Anthropic says the broader shift drove a **200% increase in PRs per engineer** at the same quality bar [^1]. Alex Albert’s framing matches that: once you can run many agents in parallel, the hard problem becomes **context management** — what matters, where agents are blocked, and when they need you — not raw generation speed [^2]. ## ⚡ TRY THIS - **Set up repo context once, then protect it.** In a new repo: run `claude`, then `/init`, refine and commit `CLAUDE.md`, add `/mcp`, define `/agents`, tune `/permissions`, and finish with `/doctor`. During normal work, compact early with `/compact`, use `/btw` for side questions, and split alternate paths into `/branch` sessions instead of polluting the main thread — a strong default from the AI For Developers guide [^3]. - **Run big changes in a plan/review loop, not a free-for-all.** Switch to `/plan`, approve the approach, execute, inspect with `/diff` as you go, then before commit run `/review` and `/security-review`. Also pick the model at the start — Sonnet for routine work, Opus for harder problems — because the guide warns that mid-task model switches and waiting too long to compact both degrade output [^3]. - **Wire repo events into async work.** Boris Cherny’s keynote demo is the clearest reproducible pattern today: configure a routine once to listen for webhooks, API calls, or a schedule; let it start Claude Code locally or on remote compute; then pair it with verification so the agent checks the fix in-browser before declaring success. The target state is simple: wake up to merge-ready PRs instead of manually starting every session [^1]. - **Use one agent as a cross-system investigator before opening a multi-day handoff.** Alex Albert says he now opens a Claude Code session with access to product databases, logs, and Slack and asks feature questions directly, instead of waiting days for a separate investigation. He also sends Claude to inspect codebases and return scope estimates like whether a feature is just a small code change plus a flag flip [^2]. ## 📡 WHAT SHIPPED - **Claude Code expanded from chat UI to agent ops stack** in the [Code with Claude 2026 keynote](https://www.youtube.com/watch?v=wjvESxKgqaQ): **Routines** for webhook/API/scheduled triggers, **Auto Fix** for code review comments, security comments, CI flakes, and merge conflicts, **Remote Control** in the iOS/Android cloud apps, **Shift Code Review** as a team of bug-finding agents, and **Cloud Security** for overnight scanning plus auto-remediation [^1]. Adoption signal is real: Anthropic says +200% PRs per engineer at the same quality bar; Shopify is using Claude Code across engineering and non-engineering teams; Mercado Libre says everyone in its 23k-engineer org runs on it, with **500k+ PRs reviewed** and **9k+ apps modernized**, targeting 90% autonomous coding [^1]. - **Claude Code’s control plane is getting more explicit.** The [slash commands guide](https://aifordevelopers.substack.com/p/claude-code-slash-commands-guide) notes that `/branch` replaced `/fork` in **v2.1.77**; `/agents` gives you delegated subagents with isolated context; `/mcp` manages external server connections; and skills in `.claude/skills//SKILL.md` can gate tool access with `allowed-tools` and `disable-model-invocation` [^3]. - **DS4 Agent is a local-first coding-agent project worth watching.** Salvatore Sanfilippo says he is building a DeepSeek 4 Flash agent that runs via CLI with the model loaded in memory, no API layer, persisted KV cache, and custom tools built around the model’s native tool-use training. The standout design detail is edit reliability: the read tool returns 4-character base64 checksum tags for file segments, and edits can target a line by checksum so replacements only happen on match; he says the current agent already calls tools successfully but still underdelivers relative to the model’s raw strength [^4]. ## 🎬 GO DEEPER - **44:01-44:39 — Overnight issue to PR via routines.** Best short clip today for seeing what async coding actually looks like: a GitHub issue lands, a repo watcher grabs it, and the work is kicked off without a human opening a fresh session [^1]. [![Code with Claude 2026: Opening Keynote](https://img.youtube.com/vi/wjvESxKgqaQ/hqdefault.jpg)](https://youtube.com/watch?v=wjvESxKgqaQ&t=2641) *Code with Claude 2026: Opening Keynote (44:01)* - **42:30-43:41 — Verification is the unlock for async coding.** The agent traces a race condition, fixes it, and verifies the behavior in-browser before calling the task done. This is the practical reason Anthropic keeps emphasizing verification instead of just more autonomy [^1]. [![Code with Claude 2026: Opening Keynote](https://img.youtube.com/vi/wjvESxKgqaQ/hqdefault.jpg)](https://youtube.com/watch?v=wjvESxKgqaQ&t=2550) *Code with Claude 2026: Opening Keynote (42:30)* - **21:36-24:23 — Salvatore on local-agent tool design.** Watch this if you care about local coding agents: he explains why keeping the model in memory, persisting KV cache, and editing by checksum could save context and reduce bad rewrites [^4]. [![Ha davvero senso oggi l'inferenza locale?](https://img.youtube.com/vi/IoE3Hi2zpwk/hqdefault.jpg)](https://youtube.com/watch?v=IoE3Hi2zpwk&t=1296) *Ha davvero senso oggi l'inferenza locale? (21:36)* - **Study guide — [Claude Code Slash Commands Guide](https://aifordevelopers.substack.com/p/claude-code-slash-commands-guide).** Best compact reference in today’s set for session controls, skills, subagents, and MCP wiring before you invent your own command soup [^3]. - **Setup thread — [Nick Baumann’s connected-device Codex workflow](https://x.com/nickbaumann_/status/2055066537002725393).** Good pattern if you want an always-on agent reachable from phone, laptop, and desktop: Mac mini as the always-connected box, connected-device thread handoff, and mutual SSH for files [^5]. *Editorial take: the real edge now is the infrastructure around the agent — persistent repo context, isolated subagents, verification, and PR-loop automation — not squeezing one more clever prompt into a long chat* [^3][^1]. --- ### Sources [^1]: [Code with Claude 2026: Opening Keynote](https://www.youtube.com/watch?v=wjvESxKgqaQ) [^2]: [Inside How Anthropic Is Building the Next Claude | Alex Albert](https://www.youtube.com/watch?v=T4ieZPIEmd8) [^3]: [Claude Code Slash Commands Guide](https://aifordevelopers.substack.com/p/claude-code-slash-commands-guide) [^4]: [Ha davvero senso oggi l'inferenza locale?](https://www.youtube.com/watch?v=IoE3Hi2zpwk) [^5]: [𝕏 post by @nickbaumann_](https://x.com/nickbaumann_/status/2055066537002725393)