Five Techniques for Reducing AI Agent Token Spend

1. Implement Prompt Caching for Repeated Context

Claude's prompt caching feature stores large system prompts and context blocks, charging only 10% of the cache creation cost for cache hits. If your agent loads the same 10KB system prompt 100 times monthly, you're reusing ~1M tokens at 90% discount.

Enable caching by setting `cache_control: {"type": "ephemeral"}` on your system prompt. This works best for agents with stable instructions, lengthy examples, or documentation that doesn't change per-request.

2. Batch Non-Urgent Requests with the Batch API

The Batch API processes requests at 50% discount if you can wait 1 minute to 24 hours for results. For async workflows—report generation, data enrichment, background analysis—batching is nearly free savings.

Group 100 requests into a single batch job instead of making individual API calls. You'll pay ~3-4K tokens per 1000-token job instead of the standard rate.

const batchRequests = [
  {
    custom_id: "task-1",
    params: {
      model: "claude-3-5-sonnet-20241022",
      max_tokens: 256,
      messages: [{role: "user", content: "Analyze this data..."}]
    }
  }
];

const batch = await client.beta.messages.batches.create({
  requests: batchRequests
});

3. Use Vision Token Optimization for Image Inputs

Images in Claude consume tokens proportional to their resolution and format. A 1080p image costs ~770 tokens; resize to 512p and you're at ~168 tokens—an 78% reduction with minimal quality loss for most agent tasks.

For screenshots, invoices, or diagrams, compress to max 768px width and JPEG quality 75-80 before sending. Your agent still reads content accurately while cutting image token spend dramatically.

4. Limit Output Length with Max Tokens Parameter

Every token generated costs you money. Set `max_tokens` to the minimum your agent actually needs—if you're extracting structured data, cap output at 500 tokens instead of defaulting to 4096. This prevents verbose rambling and forces concise responses.

Pair this with stricter prompts: instead of "explain thoroughly," ask for "a 1-sentence summary and 3 bullet points."

5. Route Requests by Complexity with Smart Model Selection

Claude 3.5 Haiku handles 80% of lightweight tasks (classification, extraction, formatting) at 1/3 the cost of Sonnet. Route simple queries to Haiku, reserve Sonnet for complex reasoning, research, or code generation.

Implement a router that checks request complexity (token count, instruction length) and selects the cheapest model that meets quality thresholds. This hybrid approach cuts spend 30-40% across your agent fleet.

Open-Source Implementation: Pantheon

The Pantheon framework (github.com/lewisallena17/pantheon) provides production-ready Next.js + Supabase templates with token optimization built in. It includes prompt caching middleware, batch job orchestration, and per-agent cost tracking.

Deploy a multi-agent system with cost monitoring in minutes—Pantheon handles caching headers, request batching, and expense dashboards so you focus on agent logic, not infrastructure.

Open-source implementation

Everything in this article runs in pantheon — a production-ready Next.js + Supabase + Claude starter. Clone it, deploy to Vercel, run PM2. The dashboard auto-commits every agent edit and reverts itself if TypeScript breaks.