Skip to main content
Tips

practitioner tip

Four Agentic AI Patterns You Already Use in Claude Code

AIErudit angle: Anthropic's five agent workflow patterns aren't theory — they're the architecture decisions you make every time you design an AI delivery workflow. This tip maps each pattern (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) to a concrete Claude Code primitive so you pick the right building block on purpose. Knowing the map turns ad-hoc prompting into repeatable, reviewable delivery.

agent-patternsclaude-codetipsworkflow-design
Tip visual

Published: Feb 24, 2026 Updated: 2026-05-10

Original: https://wmedia.es/en/tips/claude-code-agentic-ai-five-patterns

In December 2024, Anthropic published Building Effective Agents — five workflow patterns for LLM-based systems. What's not obvious: Claude Code already implements four of them natively. The fifth requires a design decision.

I've been studying and applying these patterns as part of my professional practice in Agentic AI. This map shows how each one connects to Claude Code's primitives.

Result:

Anthropic Pattern           → Claude Code Primitive
─────────────────────────────────────────────────────
Prompt Chaining             → Plan mode + Skills
Routing                     → Conditional CLAUDE.md
Parallelization             → Sub-agents / Agent Teams
Orchestrator-Workers        → Task tool (sub-agents)
Evaluator-Optimizer         → Inline skill (/evaluate)

The 5 patterns

1. Prompt Chaining — Sequential steps

Decompose a task into sequential steps where each LLM processes the previous one's output. In Claude Code: plan mode to design the sequence, skills to execute each step as a slash command.

2. Routing — Direct to the right path

Classify the input and route to the appropriate handler. In Claude Code: conditional rules in CLAUDE.md that activate behaviors based on project context.

3. Parallelization — Divide and conquer

Execute sub-tasks simultaneously. In Claude Code: sub-agents working in parallel with isolated contexts, or agent teams when teammates need to talk to each other and challenge each other's hypotheses.

4. Orchestrator-Workers — One director, many executors

A central agent that delegates, coordinates, and synthesizes. In Claude Code: the Task tool launches specialized sub-agents and collects their results. You see this every time Claude delegates a search to an Explore sub-agent.

5. Evaluator-Optimizer — Generate, evaluate, iterate

One LLM generates, another evaluates and provides feedback in a loop. Claude Code doesn't have this as a native automatic loop — the solution: an inline skill that acts as an evidence-based auditor. The hardest pattern to place, and the one that adds the most value in complex tasks.

Quick reference

PatternClaude Code PrimitiveWhen to use
Prompt ChainingPlan mode + SkillsTasks with clear phases (research → design → implement)
RoutingConditional CLAUDE.mdMulti-stack projects or context-dependent rules
ParallelizationSub-agents / Agent TeamsIndependent tasks that don't share state
Orchestrator-WorkersTask toolComplex tasks requiring central coordination
Evaluator-OptimizerInline skillValidate plans, reports, or critical implementations

The key is knowing Claude Code's 6 extension points and matching each pattern to the right one. These patterns aren't theoretical — they're architecture decisions you make every time you design a workflow.

Official docs: Building Effective Agents — Anthropic · Skills · Sub-agents