cc-agentic-patterns

v2026.09.25

Select and wire an agentic design pattern (reflection, prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, ReAct, blackboard) into the 5-layer Claude Code stack. Use this skill whenever deciding how to structure a multi-step task, whether to spawn subagents, how to run parallel review, or when to use which pattern. Triggers on: "which pattern", "orchestrate", "parallel review", "self-review", "chain of thought", "eval-optimizer loop", "blackboard", "ReAct", "how to decompose this task".

GitHub
Install command
npx skhub add thelobbi/cc-agentic-patterns
Markdown
SKILL.md

Agentic Patterns

13 patterns from Anthropic's workflow/agent taxonomy + Andrew Ng's four foundational patterns. This skill is the router; templates live in MCP cc_kb_pattern_template(name) and the pattern-router agent picks for you.

Choose by task shape

How many distinct phases?
│
├─ 1 (single generation) ──────────────► REFLECTION
│
├─ 2-4 (sequential, each uses prior) ──► PROMPT CHAINING
│
├─ 2-4 (independent, same question) ───► PARALLELIZATION
│
├─ Branching on input type ────────────► ROUTING
│
├─ Quality must meet a rubric ─────────► EVALUATOR-OPTIMIZER
│
├─ Emergent subtasks, dynamic ─────────► ORCHESTRATOR-WORKERS
│
├─ Open debug / root-cause ────────────► ReAct (thought→action→observation)
│
├─ Multi-agent collaborative review ───► BLACKBOARD (council)
│
└─ Unsure? ────────────────────────────► use agents/cc-pattern-router

Pattern quick-reference

PatternCostBest forTemplate
Reflection1.1×Default quality gate on single shotscc_kb_pattern_template("reflection")
Prompt chaining1.5–2×Multi-phase features (analyze→plan→implement→verify)cc_kb_pattern_template("prompt-chaining")
Routing1×Task-type decides the branchcc_kb_pattern_template("routing")
Parallelization1–3×Independent parallel subtasks, fan-in synthesiscc_kb_pattern_template("parallelization")
Evaluator-optimizer1.5–3×Quality-critical artifacts (configs, security code)cc_kb_pattern_template("eval-optimizer")
Orchestrator-workers2–5×Large features where subtasks emerge during workcc_kb_pattern_template("orchestrator-workers")
ReAct1.2–2×Open-ended debugging, explorationcc_kb_pattern_template("react")
Blackboard3–6×Multi-agent collaborative review/synthesiscc_kb_pattern_template("blackboard")

Wiring patterns into the 5-layer stack

Each pattern deploys artifacts across layers. Example for reflection:

LayerArtifact
L1 — CLAUDE.mdAdd rule: "Review your own output before presenting. Check for off-by-one errors, unhandled edge cases, and inconsistency with existing code patterns."
L2 — SkillNone required — behavior driven from CLAUDE.md
L3 — HookOptional: PostToolUse on Write
L4 — Agentagents/cc-audit-reviewer for explicit second-pass when risk is high
L5 — MemorySave rejection reasons to engram for learning

Fetch a specific pattern's full wiring via cc_kb_pattern_template(name).

Pattern selection via router agent

When the task is ambiguous, launch pattern-router:

Agent({
  subagent_type: "pattern-router",
  prompt: "Task: {...}. Select the optimal agentic pattern with rationale."
})

Router returns: pattern name + 5-layer wiring + cost estimate + anti-patterns.

MCP delegation

NeedTool
Fetch pattern templatecc_kb_pattern_template(name)
Decide orchestration cost capcc_docs_model_recommend(task, budget)
Recommend topology for multi-agentcc_docs_team_topology_recommend(task, complexity, team_size)

Parallelization vs Blackboard — when each applies

Both fan out to multiple agents. The difference is whether agents read each other's work:

ParallelizationBlackboard
Agents see peers?No — each gets only their scopeYes — each reads the shared board before writing
SubtasksFully independentOverlapping; one agent's finding informs another
ExampleRun security + perf + test reviews simultaneouslySecurity agent flags an issue; architecture agent reads it and re-evaluates their finding
Cost1–3×3–6× (extra rounds per agent)
When to chooseSubtasks have zero shared stateAgents must build on each other's output (adversarial or consensus-building)

If in doubt: start with parallelization. Upgrade to blackboard only when a second round of cross-agent reasoning is demonstrably necessary.

Anti-patterns

  • Using eval-optimizer for simple tasks → wasted loops, no quality gain.
  • Parallelization when subtasks share state → merge conflicts and hallucinated consistency.
  • Orchestrator-workers for fixed-scope work → supervisor overhead dominates.
  • ReAct with no stopping criterion → runs forever on fuzzy tasks.
  • Blackboard with fewer than 3 agents → not actually multi-agent; just a chain with extra steps.
Discovery
Tags

No tags published for this skill.

Version
Latest version metadata

Version

v2026.09.25

Published

Sep 25, 2026

Category

Uncategorized

License

MIT

Source path

plugins/claude-code-expert/skills/cc-agentic-patterns

Default branch

main

Latest commit

2f1269c

Tree SHA

629e050