Production-Ready AI Agent Architectures: Moving Beyond Single-Prompt LLM Integrations

], tools: ToolDefinition[], maxRetries = 2 ): Promise<ToolOutput[]> { const results: Map<string, ToolOutput> = new Map(); const toolMap = new Map(tools.map(t => [t.name, t]));

for (const step of plan) { const tool = toolMap.get(step.tool); if (!tool) { results.set(step.id, { success: false, error: Unknown tool: ${step.tool} }); continue; }

// Validate input against schema
const validation = tool.schema.safeParse(step.input);
if (!validation.success) {
  results.set(step.id, { success: false, error: `Invalid input: ${validation.error.message}` });
  continue;
}

let attempt = 0;
let output: ToolOutput = { success: false, error: 'Max retries exceeded' };

while (attempt <= maxRetries) {
  try {
    output = await tool.execute(validation.data);
    if (output.success) break;
  } catch (err) {
    output = { success: false, error: `Execution error: ${(err as Error).message}` };
  }
  attempt++;
  if (attempt <= maxRetries) {
    await new Promise(r => setTimeout(r, 200 * attempt)); // Exponential backoff
  }
}

results.set(step.id, output);

}

return Array.from(results.values()); }

### Step 4: Orchestrate with Bounded State Management
Context windows must be controlled. The orchestrator maintains a sliding window of relevant history, prunes completed steps, and injects only necessary context into subsequent planner calls.

```typescript
export class AgentOrchestrator {
  private contextWindow: string[] = [];
  private readonly maxContextTokens = 4000;

  constructor(private llm: LLMClient, private tools: ToolDefinition[]) {}

  async run(goal: string): Promise<ToolOutput[]> {
    // 1. Generate initial plan
    const plan = await generatePlan(goal, this.tools, this.llm);
    
    // 2. Execute with validation & retry
    const results = await executePlan(plan, this.tools);
    
    // 3. Update bounded context
    this.contextWindow.push(`Goal: ${goal}`);
    results.forEach(r => this.contextWindow.push(JSON.stringify(r)));
    
    // Prune oldest entries if token limit approached
    if (this.contextWindow.length > 10) {
      this.contextWindow = this.contextWindow.slice(-10);
    }
    
    return results;
  }
}

Architecture Rationale

• Decoupling Planning & Execution: Separates reasoning from I/O. Enables parallel tool execution, deterministic validation, and independent scaling of planning vs. action layers.
• Schema-First Tool Contracts: Prevents LLM hallucination of parameters, reduces injection risk, and enables static validation before runtime.
• Bounded Context Window: Pruning and explicit state tracking prevent token blowout and context drift, which are the primary causes of agent degradation over long sessions.
• Retry with Backoff & Fallback: Transient failures (rate limits, network timeouts, temporary service outages) are isolated from logic failures. The executor distinguishes between retryable and terminal errors.

Pitfall Guide

1. Unbounded Context Accumulation Appending every tool output and LLM response to the prompt causes context window exhaustion, increased latency, and cost multiplication. Best practice: implement explicit state management with token-aware pruning. Keep only task-relevant history and summarize completed steps.

2. Vague or Missing Tool Schemas LLMs will guess parameter types, omit required fields, or pass malformed JSON when tools lack strict contracts. Best practice: use Zod/Pydantic schemas, validate inputs before execution, and return structured error messages that the planner can consume for self-correction.

3. Synchronous Blocking Loops Running tool calls sequentially when they are independent creates artificial latency. Best practice: identify dependency graphs in the plan, execute independent steps in parallel using Promise.all, and resolve dependencies before downstream steps.

4. No Circuit Breaker or Fallback Strategy External APIs and LLM endpoints fail. Without circuit breakers, agents cascade into repeated failures. Best practice: implement timeout thresholds, track failure rates per tool, open circuits after N consecutive failures, and route to fallback tools or degrade gracefully.

5. Over-Engineering with Multi-Agent Systems Introducing specialized agents (researcher, writer, critic) adds coordination overhead, token cost, and debugging complexity. Best practice: start with Planner-Executor. Only introduce multi-agent patterns when tasks require fundamentally different expertise, security boundaries, or parallel workstreams that cannot be abstracted into tools.

6. Ignoring Evaluation Metrics Shipping agents without measuring reliability, cost, and latency per task leads to silent degradation. Best practice: instrument every step with metrics. Track tool success rate, planner accuracy, retry frequency, and token consumption. Run regression evals before deployment.

7. Prompt Injection via Tool Outputs Unsanitized tool responses injected back into the LLM context can trigger prompt injection or logic hijacking. Best practice: sanitize tool outputs, wrap them in explicit delimiters, and use system-level instructions that forbid interpreting tool data as commands.

Production Bundle

Action Checklist

• Define strict Zod/Pydantic schemas for every tool before implementation
• Decouple planning and execution layers to enable parallelization and independent scaling
• Implement bounded context management with token-aware pruning and state summarization
• Add retry logic with exponential backoff and distinguish retryable vs terminal errors
• Instrument execution metrics: success rate, latency, token cost, and fallback frequency
• Sanitize all tool outputs before injecting them back into the LLM context
• Run regression evaluation harnesses against known edge cases before production rollout

Decision Matrix

Configuration Template

// agent.config.ts
import { z } from 'zod';

export const agentConfig = {
  llm: {
    provider: 'openai',
    model: 'gpt-4o-mini',
    temperature: 0.1,
    maxTokens: 2048,
    timeout: 15000
  },
  planning: {
    maxSteps: 8,
    maxRetries: 2,
    contextLimit: 4000,
    pruningStrategy: 'fifo' // 'fifo' | 'semantic' | 'recent'
  },
  execution: {
    parallelEnabled: true,
    circuitBreaker: {
      threshold: 3,
      resetTimeout: 30000
    },
    fallbackTool: 'search_database_fallback'
  },
  tools: [
    {
      name: 'search_database',
      schema: z.object({
        query: z.string().min(3),
        limit: z.number().int().min(1).max(50).default(10)
      }),
      timeout: 5000
    },
    {
      name: 'generate_report',
      schema: z.object({
        title: z.string(),
        sections: z.array(z.string()),
        format: z.enum(['markdown', 'json']).default('markdown')
      }),
      timeout: 10000
    }
  ],
  observability: {
    metrics: ['latency', 'token_cost', 'tool_success_rate', 'retry_count'],
    logLevel: 'info'
  }
} as const;

Quick Start Guide

1. Initialize project: npm init -y && npm install zod openai @anthropic-ai/sdk
2. Create config file: Copy the Configuration Template into agent.config.ts and adjust provider/model settings.
3. Implement tools: Define tool schemas and execution functions matching the template structure. Ensure all inputs are validated before runtime.
4. Launch orchestrator: Instantiate AgentOrchestrator with your LLM client and tools, call run() with a goal string, and monitor metrics via your observability stack.
5. Validate: Run against 10–20 known test cases. Check tool success rate, latency, and token cost. Adjust maxRetries and contextLimit based on results before production deployment.