// CLAUDE AGENTS · ARCHITECTURAL PATTERNS

Claude Agent Design Patterns: Production System Architecture

Analyze the five architectural patterns that distinguish resilient enterprise agent systems from fragile demo scripts: Evaluator-Optimizer loops, deterministic state machines, routing trees, and checkpointed memory.

1. Canonical Pattern 1: The Evaluator-Optimizer

In this pattern, one model generates an initial implementation while a separate critic model evaluates it against strict criteria, feeding actionable critique back until quality criteria are satisfied:

flowchart LR
    Task[Task Definition] --> Gen[Generator Agent: Sonnet 5.5]
    Gen --> Artifact[Candidate Artifact]
    Artifact --> Eval[Evaluator / Critic: Sonnet 5.5]
    Eval -- Feedback & Score < 95% --> Gen
    Eval -- Score >= 95% --> Verified[Production Deployment]

2. Canonical Pattern 2: The Orchestrator-Workers

The orchestrator analyzes a high-level task, plans a dependency graph of subtasks, delegates them to specialized workers, and synthesizes the outputs into a coherent result:

Step 1 · Planning

Task Graph Analysis

Orchestrator breaks complex user goals into non-overlapping subtasks with explicit inputs and outputs.

Step 2 · Fan-Out

Parallel Worker Dispatch

Subtasks are dispatched concurrently to worker agents equipped only with relevant tool schemas.

Step 3 · Synthesis

Consolidation & Verification

Orchestrator aggregates worker outputs, checks for contradictions, and delivers unified synthesis.

3. Tutorial: Implementing an Evaluator-Optimizer in Python

evaluator_optimizer.py

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

def optimize_code(task: str, max_iterations=3):
    current_code = ""
    feedback = "Initial generation"

    for i in range(max_iterations):
        print(f"--- Iteration {i+1} ---")
        # 1. Generator
        gen_prompt = f"Task: {task}\nPrevious Code: {current_code}\nCritique to fix: {feedback}"
        gen_resp = client.messages.create(
            model="claude-3-7-sonnet-20250219",
            max_tokens=2048,
            messages=[{"role": "user", "content": gen_prompt}]
        )
        current_code = gen_resp.content[0].text

        # 2. Evaluator
        eval_resp = client.messages.create(
            model="claude-3-7-sonnet-20250219",
            max_tokens=1024,
            messages=[{"role": "user", "content": f"Critique this code for edge-case bugs and performance. Rate PASS or FAIL with specific feedback:\n\n{current_code}"}]
        )
        critique = eval_resp.content[0].text
        print(f"Critique: {critique[:150]}...\n")

        if "PASS" in critique.upper():
            print("Evaluator passed the candidate code!")
            return current_code
        feedback = critique

    return current_code

if __name__ == "__main__":
    optimize_code("Write an asynchronous LRU cache with TTL expiration in Python.")

1. Canonical Pattern 1: The Evaluator-Optimizer

2. Canonical Pattern 2: The Orchestrator-Workers

3. Tutorial: Implementing an Evaluator-Optimizer in Python

Deploy Production Claude Agent Systems