Signing you in...

Please wait while we verify your authentication

Community newsletter

Agentic Coding

Top tech stories today across software, hardware, AI, and product launches. Senior engineer audience — skip rumour churn and pre-announcement leaks. Lead with shipping-now stories and what changed for builders.

By Marius BongartsTech10 editions
Editions1 / 10
Generated by AI overnight from public sources, refreshed daily.
Agentic Coding
Saturday, May 23, 2026
AI Agents - Agentic Coding

Garry Tan ships 23-role agent team, four-agent setups get battle-tested, skills anatomy explained

1 min read

GStack drops 23 agent roles

Y Combinator's president just open-sourced his entire Claude Code setup.

Garry Tan released gstack, a framework that transforms Claude Code into a virtual engineering team with 23 specialized roles—CEO, designer, eng manager, QA—and 40+ slash commands like /office-hours for product discovery and /qa for live browser testing with real Chromium [Source: GitHub]. He claims 810× his 2013 productivity pace, shipping 3 production services and 40+ features in 60 days while running YC full-time. The framework includes GBrain for persistent cross-session memory and works with 10 other AI coding agents including Cursor.

Free, MIT-licensed, and worth stealing the sprint workflow structure even if you don't adopt the whole thing.

Four-agent team structure

The magic number might be four, not five.

A detailed workflow breakdown from Zyte shows how to structure a team of specialized agents with strict permission boundaries: an architect agent for read-only planning, a scout for codebase exploration, a coder for execution, and a reviewer for quality checks [Source: Zyte]. The read-only constraint on architect and scout prevents accidental damage—a pattern that maps directly to the 3-5 agent ceiling Anthropic documented in yesterday's issue.

The real insight: restart sessions with fresh context instead of letting the window saturate.

How skills actually load

Skills aren't magic—they're progressive disclosure.

A deep dive from Firecrawl explains the mechanics: Claude scans available skills at startup reading only name and description, then loads full instructions when a task matches [Source: Firecrawl]. Two types matter: Capability Uplift skills give Claude new abilities it lacks, while Encoded Preference skills lock in specific workflows so Claude stops defaulting to generic patterns. The article highlights Andrej Karpathy's Guidelines skill—think before coding, simplicity first, surgical changes—as a way to prevent over-engineering.

Convert your repeated prompts into skills or keep teaching Claude the same lessons every session.

Sources
My agentic coding setup: Claude Code, multi-agent orchestration ...
My agentic coding setup: Claude Code, multi-agent orchestration ...
21 hours ago ... What I have now is a battle-tested setup that survived that process, not because the tools are exceptional in isolation, but because I made deliberate choices ...
zyte.com
AI Summary

The author shares a battle-tested agentic coding setup combining Claude Code CLI, Cursor (via Codex), and OpenRouter for experimentation. Key workflow optimizations include: using a shell function to open a fully configured workspace in one command, enforcing plan mode before any file editing, maintaining a CLAUDE.md file at project root and globally to preserve context and conventions between sessions, and implementing a four-agent team with specialized roles (architect for planning read-only, scout for codebase exploration, coder for execution, reviewer for quality checks). The setup prioritizes permission constraints—read-only agents prevent accidental damage—and uses pay-per-usage pricing to maintain cost discipline. Critical habits include defining concrete metrics before scraping tasks, restarting sessions with fresh context rather than letting the context window saturate, and converting repeated prompts into reusable skills. The author emphasizes that domain expertise matters more than tool choice—your accumulated judgment as a domain expert is the actual advantage, and your setup should amplify your existing strengths rather than copy someone else's wholesale.

Visit source
GitHub - garrytan/gstack: Use Garry Tan's exact Claude Code setup
GitHub - garrytan/gstack: Use Garry Tan's exact Claude Code setup
20 hours ago ... Other AI Agents. gstack works on 10 AI coding agents, not just Claude. Setup ... Researches the codebase first, then writes reference / how-to / tutorial / ...
github.com
AI Summary

Garry Tan, President of Y Combinator, has released gstack, an open-source framework that transforms Claude Code into a virtual engineering team with 23 specialized roles and 40+ slash-command skills. The system implements a structured sprint workflow (Think → Plan → Build → Review → Test → Ship → Reflect) designed for solo founders and small teams using AI-assisted coding. Key skills include /office-hours for product discovery, /plan-ceo-review and /plan-eng-review for architectural planning, /review for code inspection, /qa for live browser testing with real Chromium, and /ship for CI/test/deploy automation. Tan demonstrates productivity gains using this system: his 2026 run rate shows approximately 810× his 2013 pace when measured in logical code changes, shipping 3 production services and 40+ features in 60 days while running YC full-time. The framework also works with 10 other AI coding agents including Cursor, OpenClaw, and OpenAI Codex through configurable host detection. Advanced features include GBrain for persistent cross-session memory, multi-agent coordination via /pair-agent, design-to-HTML pipeline with taste learning, iOS device testing, security audits (OWASP/STRIDE), and continuous checkpoint mode for crash recovery. The tool is free and MIT-licensed with optional telemetry (off by default) and comprehensive browser automation via a dedicated GStack Browser with anti-bot stealth and sidebar agent support for authenticated workflows.

Visit source
Best Claude Code Skills to Try in 2026 - Firecrawl
Best Claude Code Skills to Try in 2026 - Firecrawl
5 hours ago ... "Help me optimize this landing page for conversions" "Write homepage copy ... Building the top Claude Code skills into your workflow. Skills solve the ...
firecrawl.dev
AI Summary

Claude Code skills are directories containing a SKILL.md file that extend Claude's capabilities in specific ways. They load progressively—Claude scans available skills at startup (reading only name and description), then loads full instructions when a task matches. Two main types exist: Capability Uplift skills (like Firecrawl for web scraping) give Claude new abilities it lacks, while Encoded Preference skills (like Frontend Design or Andrej Karpathy's Guidelines) encode specific workflows so Claude executes consistently without defaulting to generic patterns. The article highlights 14 production-ready skills worth installing for agentic workflows: Firecrawl enables reliable web data access for agents, GStack turns Claude into a full AI engineering team with office hours and QA roles, Andrej Karpathy's Guidelines encode behavioral rules (think before coding, simplicity first, surgical changes, goal-driven execution) to prevent Claude from over-engineering or making silent wrong assumptions, Frontend Design pushes past generic AI aesthetics to intentional visual direction, Superpowers orchestrates multi-agent development with planning and test-driven discipline, and Vercel's Web Design Guidelines and React Best Practices automate quality gates for accessibility and performance. For debugging and refinement, the Webapp Testing skill lets Claude interact with live browser state, while Document Skills handle real file creation and manipulation. Custom skills can be built using the Skill Creator plugin to encode team-specific preferences and workflows that Claude keeps forgetting between sessions.

Visit source
Compiled overnight by MorningMail.aiDelivered at 04:55 AM