Duyetbot Agent
Advanced

Token Optimization

75% savings via routing, batching, trim. Pattern match + simple agents.

TL;DR: Hybrid classify (80% patterns=0 tokens). Batch 3-5 msgs/call. Trim history. Route simple=150t vs orch=1500t.

Table of Contents

Savings Table

Real-world: 100 queries/day.

From docs/token-optimization-guide.md.

MechanismSavingsTokens/Query
Hybrid Classify80%60 vs 300
Batching55%200/call
Simple Routing70%75 avg
Dedup10%Skip retries
Total75%7.5k vs 30k

Badge: 75% ✅

Tiers

               Query
                |
                v
         +--------------+
         |   Hybrid?    |
         +--+----------++
            |          |
      80% PATTERN   20% LLM
            |          |
            v          v
       SimpleAgent  Classify 300t
      (50-150t)      |
                     v
                 +-------------+
                 |  Complex?   |
                 +--+--------+-+
                    |        |
                   NO       YES
                    |        |
                    v        v
                Simple   Orch
                100t     1500t

Integrates batch/trim.

Trim Snippets

Trim before LLM. From history.ts.

// packages/cloudflare-agent/src/history.ts
export function trimHistory(messages: Message[], maxLength: 10): Message[] {
  return messages.length > maxLength
    ? messages.slice(-maxLength)
    : messages;
}
 
// Use: const recent = trimHistory(state.messages, 10);

Combine batch: batch-types.ts join('\n---\n').

P50 costs: $0.0225/100q.

Pattern Quiz

Q: Greeting "hi"? Tokens?

A: 0 (pattern match) ✅
B: 300 (LLM classify)
C: 1500 (orchestrator)

Measure tokens: Add tokensUsed logs. Optimize patterns!

On this page