Skills, Agents & Teams

Cutting-Edge Techniques in AI Agentic Programming and Design — v3.0

Executive Summary

Version 3.0 is a practitioner's guide, not a market survey. Built on two months of daily production operation — 190+ research projects, 21,298 passing tests, 34 custom skills, 26 custom agents, and proven orchestration patterns running across 6 concurrent branches — this report documents what actually works when you push AI agentic systems to scale. Five new sections cover hooks (4 types, 20+ lifecycle events), agent teams (multi-session coordination), production-proven orchestration patterns (Fleet Mission, Gastown Convoy, GUPP Propulsion), memory and state management, and real case studies with hard metrics. Updated throughout with Anthropic's current best practices for Claude Code skills, subagents, hooks, and the Claude Agent SDK.

34+ Custom Skills
26 Custom Agents
15 Report Sections
242+ Projects Shipped
24,336 Tests Passing
600K+ Lines Generated

Architecture Catalog

🤖

Agents

26 production agents + 33 reference

Custom subagents: GSD workflow (planners, executors, verifiers), research (fleet commander, synthesizer), DevOps (security reviewer, pipeline analyzer), and specialized workers.
⚙

Skills

34 custom + 5 bundled + 33 reference

Production skills: orchestration (Fleet Mission, Gastown, GUPP), communication (mail-async, nudge-sync), pipelines (done-retirement, refinery-merge), plus DevOps and workflow skills.
👥

Teams

12 team compositions + patterns

Multi-agent teams: code review, DevOps pipeline, incident response, research fleet, and the Gastown convoy model for 50+ project orchestration.
📑

Report

15 sections + references

Comprehensive analysis from fundamentals through production case studies. 5 new sections on hooks, agent teams, orchestration patterns, memory, and real-world results at scale.

What's New in v3.0

5 New Report Sections

Section 11: Hooks — 4 hook types across 20+ lifecycle events with production patterns.
Section 12: Agent Teams — Multi-session coordination with shared tasks and messaging.
Section 13: Orchestration Patterns — Fleet Mission, Gastown Convoy, GUPP Propulsion — proven at 50+ project scale.
Section 14: Memory & State — Auto memory, CLAUDE.md hierarchy, context handoffs, crash recovery.
Section 15: Case Studies — 242+ projects, NASA catalog, Artemis II live coverage, GPU math co-processor.

Updated Throughout

Skills: New frontmatter (paths, effort, model, hooks, agent, context: fork), bundled skills (/batch, /simplify, /debug), shell injection, string substitutions.
Subagents: Worktree isolation, persistent memory, MCP scoping, skills preloading, model routing, /agents command.
Hooks: HTTP hooks, prompt hooks, agent hooks, if-conditions, 20+ events.
Current as of: Claude Code v2.1.88, Claude Opus 4.6, April 2026.

Report Sections

1

Fundamentals

Core concepts, MCP, Agent SDK, and the Claude Code platform.

 2

Skills

SKILL.md format, frontmatter, activation, bundled skills.

 3

Subagents

Custom agents, model routing, worktrees, memory.

 4

Token Optimization

1M context strategies, caching, model routing.

 5

Performance

Extended thinking, effort levels, scaling.

 6

Applications

Use cases, enterprise strategy, implementation.

 7

Accuracy

Guardrails, safety, self-modifying system security.

 8

Future Trends

2026 H2 outlook, Agent SDK, plugins ecosystem.

 9

DevOps Lifecycle

8-phase DevOps with AI agents, CI/CD, AIOps.

 10

GSD Orchestrator

57 commands, wave execution, lifecycle management.

 11

Hooks NEW

4 types, 20+ events, production automation.

 12

Agent Teams NEW

Multi-session coordination, shared tasks.

 13

Orchestration NEW

Fleet Mission, Gastown, GUPP — proven at scale.

 14

Memory & State NEW

Auto memory, handoffs, crash recovery.

 15

Case Studies NEW

242+ projects, real metrics, lessons learned.