--- title: "Claude Code: swarm mode (agent teams)" url: https://perrotta.dev/2026/02/claude-code-swarm-mode-agent-teams/ last_updated: 2026-02-09 --- Goodness. Who asked for a feature to burn tokens? [Orchestrate teams of Claude Code sessions](https://code.claude.com/docs/en/agent-teams): > Coordinate multiple Claude Code instances working together as a team, with > shared tasks, inter-agent messaging, and centralized management. Demo running in the git repository of this blog: ``` ▐▛███▜▌ Claude Code v2.1.37 ▝▜█████▛▘ Opus 4.6 · Claude Team ▘▘ ▝▝ ~/Workspace/perrotta.dev ❯ swarm mode to analyze my blog ⏺ I'll set up a team of agents to analyze your blog from multiple angles in parallel. ⏺ Now let me spawn 4 agents to work in parallel on each analysis task. ⏺ 4 agents launched (ctrl+o to expand) ├─ @content-analyst (Explore) │ ⎿ Analyze blog content and trends (shift+↑ to manage) ├─ @tech-analyst (Explore) │ ⎿ Analyze Hugo technical setup (shift+↑ to manage) ├─ @structure-analyst (Explore) │ ⎿ Analyze site structure and linking (shift+↑ to manage) └─ @quality-analyst (Explore) ⎿ Analyze code quality and practices (shift+↑ to manage) ⏺ 4 analysis agents are now running in parallel: ┌───────────────────┬─────────────────────────────────────────┬─────────┐ │ Agent │ Focus Area │ Status │ ├───────────────────┼─────────────────────────────────────────┼─────────┤ │ content-analyst │ Posts, topics, tags, writing trends │ Running │ ├───────────────────┼─────────────────────────────────────────┼─────────┤ │ tech-analyst │ Hugo config, layouts, shortcodes, theme │ Running │ ├───────────────────┼─────────────────────────────────────────┼─────────┤ │ structure-analyst │ Internal linking, assets, navigation │ Running │ ├───────────────────┼─────────────────────────────────────────┼─────────┤ │ quality-analyst │ Templates, CSS/JS, code quality │ Running │ └───────────────────┴─────────────────────────────────────────┴─────────┘ I'll compile the results as they come in. ✢ Analyzing Hugo technical setup… (2m 0s · 0 tokens) ╒═ team-lead · shift + ↑/↓ to select ├─ @content-analyst: Running Get recent and oldest posts with metadata… · 19 tool uses · 34.3k tokens ├─ @quality-analyst: Reading 5 files… · 24 tool uses · 70.6k tokens ├─ @structure-analyst: Reading 5 files… · 26 tool uses · 38.5k tokens └─ @tech-analyst: Searching for 2 patterns, reading 3 files… · 73 tool uses · 95.9k tokens ``` **Lesson learned**: start Claude in [YOLO](https://simonwillison.net/2025/Oct/22/living-dangerously-with-claude/) mode: ```shell % which claudey claudey: aliased to claude --dangerously-skip-permissions ``` Some folks are starting to suggest to alias this invocation to `cc`, but I find it to be poor taste. `cc` is the C compiler. ...otherwise you'll be nagged to death with permissions prompts from multiple agents. [Decision fatigue](https://en.wikipedia.org/wiki/Decision_fatigue) is real, folks. ``` @tech-analyst❯ Complete Hugo technical setup analysis ⏺ Great, the tech analyst has finished. Let me update its task and wait for the others. ⏺ Excellent analysis from the tech analyst. Waiting on the other 3 agents to finish their analysis... ⏺ The tech analyst is idle now — its work is already captured. Waiting on content, structure, and quality analysts to report back. ⏺ Still waiting on the other 3 agents (content, structure, quality) to complete their analysis... @quality-analyst❯ Complete code quality analysis report ⏺ Quality analyst is done too. Let me update its task. ⏺ 2 of 4 analyses complete. Still waiting on content-analyst and structure-analyst... ⏺ Waiting on content-analyst and structure-analyst to finish... ⏺ Still waiting on the last 2 agents... ⏺ Waiting for content-analyst and structure-analyst to complete their reports. @content-analyst❯ Complete blog content analysis: 1041 posts, trends, tags, topics ⏺ Content analyst is done with a thorough report. Let me update its task. ⏺ 3 of 4 complete. Just waiting on structure-analyst now... ⏺ Waiting on structure-analyst to finish the last report... ⏺ Still waiting on the structure-analyst... @structure-analyst❯ Complete site structure analysis report ⏺ All 4 agents have reported back! Let me update the last task and shut down the team. ✻ Baked for 5m 17s ⏺ --- Blog Analysis: perrotta.dev Here's the consolidated analysis from all 4 agents. At a Glance ┌──────────────────────┬─────────────────────┐ │ Metric │ Value │ ├──────────────────────┼─────────────────────┤ │ Total posts │ 1,041 │ ├──────────────────────┼─────────────────────┤ │ First post │ Sep 2014 │ ├──────────────────────┼─────────────────────┤ │ Unique tags │ 27 │ ├──────────────────────┼─────────────────────┤ │ Top tag │ dev (902 posts) │ ├──────────────────────┼─────────────────────┤ │ Median post length │ 44 lines │ ├──────────────────────┼─────────────────────┤ │ Internal cross-links │ 207 (bidirectional) │ ├──────────────────────┼─────────────────────┤ │ Graph nodes/edges │ 1,039 / 213 │ ├──────────────────────┼─────────────────────┤ │ Template files │ 33 │ ├──────────────────────┼─────────────────────┤ │ Pre-commit hooks │ ~40 │ ├──────────────────────┼─────────────────────┤ │ External theme │ None (fully custom) │ └──────────────────────┴─────────────────────┘ --- Content Profile Growth trajectory is dramatic — from 1 post in 2014 and a 5-year hiatus (2017-2021) to 598 posts in 2025 and 153 already in Feb 2026. The coding section (LeetCode, ByteByteGo, AoC) drives the volume surge, with batch publishing days hitting 43 posts in a single day. Two distinct voices: 87% of posts are tagged dev (technical), while 19% are tagged serenity (philosophical, short reflections — some just 9 lines). The bestof tag (148 posts, 14%) provides a curated recommendation layer. Top topics: LeetCode solutions (201), serenity reflections (194), bestof picks (148), AI/LLM (50), git (48), privacy (43), macOS (41), Kubernetes (35). Writing style: Overwhelmingly short-form (median 44 lines, longest 367). Text-only — only 3 posts contain images. --- Technical Architecture Fully custom Hugo setup — no external theme, all 33 templates in layouts/. Catppuccin color scheme (Latte/Mocha) with modern CSS (light-dark(), view transitions, prefers-reduced-motion). All CSS inlined (~1200 lines), JS deferred and fingerprinted. Standout features for a static site: - Client-side full-text search with lazy-loaded JSON index - D3.js interactive force-directed graph (full site + per-post mini-graphs) - Precomputed backlinks, related posts, and "Previously" chains (via Python script) - Keyboard navigation (j/k, arrows, f for fullscreen) - IndieWeb microformats (h-card, h-entry, h-feed) - Multiple output formats: HTML, RSS, Markdown, OPML, llms.txt, JSON - Multilingual (EN/PT) with conditional language switcher Build pipeline is robust: Justfile with 19 recipes, Python precomputation step, and the most thorough pre-commit setup I've seen for a personal blog (40 hooks covering formatting, security, content quality, and Hugo-specific validation). --- Site Structure External links heavily favor GitHub (517), LeetCode (204), Wikipedia (84), and StackOverflow (45). Internal cross-referencing centers on the Advent of Code hub post (28 inbound links). SEO is comprehensive: JSON-LD, OpenGraph, Twitter Cards, canonical URLs, custom robots.txt blocking AI crawlers, sitemap pinging, llms.txt. Privacy-first: No analytics, no cookies, no ads, no tracking. Email-based replies. EXIF stripping on images. Warrant canary page. --- Strengths 1. Performance — Inlined CSS, deferred JS, lazy loading, no external fonts, 512KB Club green team member 2. Automation — 40 pre-commit hooks enforce quality, formatting, security, and Hugo-specific rules automatically 3. Rich features — Search, graph, backlinks, related posts, keyboard nav, fullscreen — all on a static site 4. Privacy — Zero tracking, no third-party dependencies for core functionality 5. Content curation — bestof tagging and disciplined taxonomy (only 27 tags for 1,041 posts) Potential Improvements 1. D3.js weight — d3.v7.min.js (~270KB) loads on every post page for the mini-graph; could be lazy-loaded 2. Inline styles in templates — Several partials use inline style attributes instead of CSS classes 3. Chroma CSS always inlined — ~78 syntax highlighting rules load even on pages without code blocks 4. Hugo CI version drift — Deploy uses 0.152.2, prek uses 0.140.0, local is 0.155.3 5. Portuguese content — Only 8 PT posts despite multilingual setup; could be expanded or streamlined ✻ Cooked for 36s ⏺ All 4 agents have shut down. Let me clean up the team. ⏺ 4 teammates shut down gracefully ⏺ Team cleaned up. The full analysis is in my summary above — let me know if you'd like to dig deeper into any specific area, or if you want actionable recommendations on any of the improvement areas identified. ``` **Lesson learned**: launch `tmux` before activating the swarm. You'll be able to see in multiple tmux panes at a glance all agents working simultaneously. This feature is supported out-of-the-box.