Skip to content

Roadmap — Advanced Patterns

The foundations tier teaches the disciplines. The advanced tier teaches how to automate them.

This curriculum's six phases are the foundations — the prompting, context, verification, memory, spec, and orchestration disciplines every agent-first engineer needs. The Advanced Patterns tier (below) builds on them. Its throughline:

The advanced thesis

Most advanced "tricks" — prompting patterns, workflow habits, review rituals — are best practices you should encode and automate with hooks, skills, instructions, and CI rather than perform by hand. The foundations teach the practice; the advanced tier teaches the automation and the systems around it.

flowchart LR
    F["Foundations tier (done)<br/>P1-P6"] --> A["Advanced Patterns tier (next)"]
    A --> T1["Skills & hooks<br/>deep-dives"]
    A --> T2["Advanced agent<br/>features"]
    A --> T3["Memory & living<br/>docs"]
    A --> T4["Token<br/>optimization"]
    A --> T5["Security &<br/>governance"]
    A --> T6["Cloud / voice /<br/>mobile"]
    A --> T7["Evals &<br/>spec-at-scale"]

Maturity legend: 🟢 Teach-now (shipping, stable) · 🟡 Emerging (real but moving fast) · 🔴 Bleeding-edge (early/experimental). This is a living document — topics graduate into full lessons over time.

Next up — the Advanced tier (priority)

The seven we'd write first, because they're mature and highest-leverage for an agent-first codebase:

  1. Anatomy of a Skill 🟢 — ✅ shipped → Lesson 7.1. SKILL.md internals end-to-end: frontmatter, scripts//references// assets/, progressive disclosure, packaging as plugins. Skills are how you encode a repeatable procedure once and let any agent reach for it. Agent Skills spec · anthropics/skills · Cursor Skills
  2. Hooks, deep 🟢 — ✅ shipped → Lesson 7.2. Every lifecycle event and handler type; patterns for enforcing rules, blocking danger, and reducing approval fatigue. Hooks are how you make a best practice deterministic instead of hopeful. Claude Code hooks · Codex hooks · Cursor hooks
  3. MCP, deep 🟢 — ✅ shipped → Lesson 7.3. Designing Model Context Protocol servers to expose your codebase, data, and tools to any agent over one open standard. The portable way to give agents new capabilities. MCP specification · modelcontextprotocol/modelcontextprotocol
  4. Subagents & multi-agent orchestration 🟢 — ✅ shipped → Lesson 7.5. Least-privilege subagents, writer/reviewer pairs, parallel fan-out, model-per-task. Scale throughput and quality without flooding one context. Claude Code subagents · Cursor agent best practices
  5. Security, permissions & sandboxing 🟢 — ✅ shipped → Lesson 7.4. Permission models, filesystem/network isolation, and prompt-injection defense for agents that read untrusted input. Non-negotiable before autonomy. Claude Code sandboxing · OWASP Top 10 for LLM/Agentic Apps
  6. Prompt & context caching 🟢 — ✅ shipped → Lesson 8.1. Cache stable prefixes (system prompt, AGENTS.md, big files) to cut cost and latency on long-running agents. The cheapest token optimization that exists. Anthropic prompt caching
  7. Computer use & browser/UI agents 🟢 — ✅ shipped → Lesson 8.2. Vision + control so agents can test UIs, fill forms, and self-verify visually. Closes the loop for work that has no unit-test oracle. Anthropic advanced tool use · Managed Agents overview

Backlog — by theme

Theme 1 · Skills, Hooks & Plugins (extensibility)

Topic What & why Maturity Source
Plugins & marketplaces ✅ 8.3 Bundle skills + commands + hooks + agents to share across a team/org 🟢 Agent Skills spec · Claude agent skills
Skill evals Measure whether a skill triggers + performs; tune the description 🟡 anthropics/skills

Theme 2 · Advanced agent features

Topic What & why Maturity Source
Background & long-running agents Multi-hour/day tasks across sessions; the harness that makes them reliable 🟡 Effective harnesses for long-running agents
Managed agents A managed loop + sandbox to run an agent as autonomous infra 🟡 Claude Managed Agents
Plan mode, checkpoints, output styles Per-agent control surfaces and when each matters 🟢 Claude Code docs

Theme 3 · Memory, learnings & living documentation

Topic What & why Maturity Source
Structured note-taking Persist learnings outside the context window so compaction can't lose them 🟢 Effective context engineering
Anti-staleness / self-updating docs Hooks that keep AGENTS.md + docs current from real changes; ADRs as specs 🟡 Effective context engineering

Theme 4 · Token optimization (open problem)

Topic What & why Maturity Source
Prompt/context caching Reuse stable prefixes to cut cost + latency 🟢 Anthropic prompt caching
Semantic code indexing / RAG-for-code Retrieve only the relevant code/docs instead of dumping the repo 🟡 MCP specification
LLM Wiki in your codebase A curated, agent-readable knowledge layer for docs/learnings — kept fresh and token-cheap 🔴 Optional add-on — pending a colleague's open-source release; will be added here as a reference + an optional scaffold-agent-project add-on once published.

Theme 5 · Security & governance at scale

Topic What & why Maturity Source
Sandboxing & permissions Isolate fs/network; least-privilege tool access 🟢 Claude Code sandboxing
Prompt-injection defense Defend agents that read untrusted files/web/tools 🟡 OWASP Top 10 for LLM/Agentic Apps
Observability, cost control & fleets Telemetry, spend caps, and managing many agents 🟡 Claude Managed Agents

Theme 6 · Cloud, voice & mobile coding

Topic What & why Maturity Source
Cloud agents Delegate to isolated cloud VMs; merge-ready PRs with artifacts 🟢 Cursor Cloud Agents · Codex cloud
Headless / CI-driven agents agent -p in CI; issue→PR automation 🟢 GitHub Copilot coding agent
Automated PR review Agents reviewing PRs with full project context 🟢 Codex GitHub code review
Voice & mobile / ticket automation Code from a phone or by voice; turn tickets into PRs 🔴 Codex web/cloud

Theme 7 · Evals & spec-driven at scale

Topic What & why Maturity Source
Agent evaluation & benchmarking Measure autonomy/safety so agents don't degrade silently 🟡 SWE-bench
Spec-driven development at scale Specs as the audit trail + source of truth across many agents/teams 🟡 GitHub Spec Kit

How this grows

Topics graduate from this roadmap into full phases (lessons + diagrams + quiz.json) following the same authoring rubric the foundations used. Each advanced lesson ends, like the foundations, by showing what the scaffolder can generate so you automate the pattern rather than perform it by hand. Suggestions and sources welcome via issues/PRs.