Coding Agents in Practice (5/5) — Cost Management: Tokens, Caching, Routing

AI ์ฝ”๋”ฉ ์—์ด์ „ํŠธ ์‹ค์ „ (5/5) — ๋น„์šฉ ๊ด€๋ฆฌ: ํ† ํฐ·์บ์‹ฑ·๋ผ์šฐํŒ…

The answer to "why is this so expensive?" is almost always which tokens didn't hit cache.

ํ•ต์‹ฌ ์š”์•ฝ

  • Four levers: token unit price / cache hit rate / model routing / context management
  • Primary sources: Anthropic prompt caching docs, OpenAI prompt caching docs, Claude Code pricing
  • Strongest single lever: prompt caching — keeping system prompts and CLAUDE.md cache-hit can drop input cost to ~10%
  • Caveat: Anthropic cache writes are base × 1.25 (5-minute TTL) or × 2 (1-hour TTL). Low hit rate can make caching more expensive

1. Where the cost comes from

LLM token cost = (input tokens × input rate) + (output tokens × output rate), multiplied by cache state.

Token type Rate (Claude Sonnet 4.6 example) Note
Input (cache miss) Base 1.0× Sent fresh each time
Input (cache write, 5-min TTL) Base × 1.25 Cost of creating the cache entry
Input (cache write, 1-hour TTL) Base × 2.0 Costs more to live longer
Input (cache hit) Base × 0.1 ~10% on hit
Output Base × ~5 Typically 5× more than input

Key observations: - Output is more expensive than input → simply asking the model for short answers saves cost. - A cache hit is 1/10 the cost → CLAUDE.md, system prompts, and frequent context must be designed to hit cache. - Cache writes are 1.25× (5-min) or 2× (1-hour) → there's a break-even point. The 1-hour cache costs 2× to write, so it pays back only after at least 4 hits (covered in detail in series C-3).

2. Lever 1 — Prompt Caching

The strongest single lever. Cacheable items: - System prompts - CLAUDE.md / rule files - Frequently used static context (tool definitions, design docs)

Application pattern (Anthropic): - Mark cache_control in API messages. - Cache by block — the same block must be sent in the same order to hit. - 5-minute TTL is default. Claude Code is known to use 1-hour TTL by default.

Common mistakes: - The front of the cached block changes each request → no hit. Move volatile content to the back; keep the static parts up front to cache. - Using 1-hour cache for low-hit-rate work → 2× write cost never recouped. - Cached content is too long — caching cuts cost, it does not split it. Manage context length separately.

3. Lever 2 — Model Routing

You don't always need the most expensive model.

Task type Suitable model Reasoning
Deep reasoning / design / hard debugging Opus 4.7 Reasoning depth pays back via fewer retries
Standard code changes / typical PRs Sonnet 4.6 Best price-quality balance
Simple text / classification / summarization Haiku 4.5 ~1/10 the unit cost; fastest
Background monitoring / polling Haiku or local oMLX Avoids external API calls entirely

Two routing axes: - Quality threshold: "If this fails, do we redo the work?" If yes, pay for the better model. - Latency tolerance: "Does the user need a 5-second answer?" If yes, smaller model + cache.

Routing gets its own treatment in series C-2. The point here is: a single-model strategy is almost always more expensive than necessary.

4. Lever 3 — Context management

Long context costs per token. Four ways to shrink:

  • Externalize memory: rarely-referenced information goes to memory files; only the index (MEMORY.md) sits in context.
  • Summary reports: subagents return summaries instead of full output (series part 4).
  • Auto-compaction: Claude Code auto-compacts when context fills — but compaction can lose the reasoning behind decisions. Manually consolidate before critical decisions.
  • Session separation: don't pile too much into one session; start a new one when a task ends.

5. Lever 4 — Output length control

Output is ~5× more expensive than input. Ways to shorten output:

  • Explicit limits: "report under 200 words," "code only, no explanation" in system / skill prompts.
  • Structured output: JSON schemas, tables, checklists naturally compress.
  • Pick one: ask for the single best option, not three options — request comparison only when truly needed.
  • Avoid restatement: block the "summary + conclusion" double-restate pattern via the system prompt.

6. Monitoring — Without measurement, nothing shrinks

Step one of cost management is measurement.

Minimum monitoring: - Daily token usage (input / output / cache hit rate). - Cost distribution per task type (code change vs. review vs. docs). - Per-model usage share.

Tools: - Anthropic Console Usage page shows daily, weekly, monthly token breakdowns. - Claude Code's /cost slash command displays per-session cost (CLI versions). - A custom routing layer should log per-call cost.

Empirical signal: if cache hit rate < 30%, revisit cache placement (likely the front of the cache block is changing each request).

7. At a glance

Lever Magnitude Application difficulty First step
Prompt Caching Very large (1/10) Medium (block re-layout) Cache CLAUDE.md / system prompt
Model Routing Large (10× per-task swing) Medium (task-classification rules) Ask "does this task need Opus?"
Context Management Medium (grows with task length) Low Summary-reports from subagents
Output Length Medium (output costs 5× input) Low Add length limits to system prompts

Recommended starting order: 1. Apply caching to system prompts + CLAUDE.md (often 50%+ savings immediately) 2. Route Haiku for simple tasks 3. Add output-length limits 4. Start daily cache-hit monitoring

Series wrap — Part 5/5

The Coding Agents in Practice series binds five axes — workflow (1), tool comparison (2), MCP (3), multi-agent (4), cost (5). Each post supports the others: workflow has to be in place before multi-agent helps, and multi-agent has to land before cost routing pays off.

The next campaign (Series C — AI Operations Economics, 4 parts) deepens the cost discussion. Part 1 is the measurement traps in token-cost structure.

References

  • Anthropic, Prompt Caching — docs.claude.com/en/docs/build-with-claude/prompt-caching (verified 2026-05-05).
  • Anthropic, Pricing — claude.com/pricing (verified 2026-05-05).
  • OpenAI, Prompt Caching — platform.openai.com/docs/guides/prompt-caching (verified 2026-05-05).
  • Series parts 1 (workflow) and 4 (multi-agent).

This is part 5/5 — the final entry in the Coding Agents in Practice series.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »