> ## Documentation Index
> Fetch the complete documentation index at: https://agent-harness101.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Sources & notes

> Published figures, their origins, and the caveats that keep them honest.

This walkthrough is a system-design reasoning exercise, not a claim about Claude Code’s confirmed internals. Where a section shows a number, it is cited below. Where it shows a worked example, it is labelled illustrative — no invented figures are attached to it. That line matters: keep it intact if you remix this.

## Sourced figures

### Prompt caching — up to 90% cost / 85% latency reduction

Anthropic’s own published figures for long, cached prompts. Used on the token-reduction question as the one lever with an official number; the other five levers are real patterns shown without invented percentages.

[Anthropic documentation — prompt caching](https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching)

### Model-routing hook — \~90% mean token savings, 350-line threshold, 10–30s added latency

From Spotify Engineering’s “Portal / shunt” post (Sept 2026). **This is a third-party plugin**, not native Claude Code behaviour — only the `PreToolUse` hook mechanism itself is native. The real mechanism is a single binary line-count check; it does **not** detect “reasoning need.” Edits, debugging, architecture decisions and safety-critical code are excluded from routing by design, not per-call detection. Treat the 90% as that team’s result on one Java monorepo, not a universal one.

[engineering.atspotify.com — Portal / shunt](https://www.engineering.atspotify.com/2026/9/portal-by-spotify-cut-my-claude-code-token-usage-by-90)

### Multi-agent research — +90.2% task success, \~15× token cost, 80% of the gap explained by token spend

From Anthropic’s “How we built our multi-agent research system.” This is Anthropic’s **Research** system (Opus lead + Sonnet subagents on breadth-first search tasks), **not Claude Code specifically**. Anthropic explicitly notes the pattern is **worse** for tightly-coupled work like most coding — which is exactly why the following question gates delegation instead of applying it everywhere.

[anthropic.com/engineering — built-multi-agent-research-system](https://www.anthropic.com/engineering/built-multi-agent-research-system)

### Context rot — referenced qualitatively, no single stat

From Anthropic’s “Effective context engineering for AI agents.” As tokens accumulate, models measurably lose the ability to use the full window well. Left number-free on purpose — Anthropic doesn’t publish one single figure for it.

[anthropic.com/engineering — effective-context-engineering](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)

### Seven local stores — observed layout, not a stable API

Based on inspecting Claude Code’s local `~/.claude` directory. The JSONL session-log format and location are independently confirmed by community tooling; treat the other exact paths as an **observed layout** of a specific version, not a documented stable public API.

### git worktree isolation — native, documented feature

`isolation: worktree` is a documented, native Claude Code feature: a separate working directory with its own branch, sharing the parent repo’s history.

[code.claude.com/docs — worktrees](https://code.claude.com/docs/en/worktrees)

### Core tool set & planning — real, documented features

Read / Grep / Glob / LS, Edit / Write, Bash, Plan Mode and TodoWrite are real, documented Claude Code features.

[code.claude.com/docs](https://code.claude.com/docs)

## Deliberately illustrative

The following are framed as reasoning, not measured benchmarks — keep that framing if you reuse them:

* **The two-task worked example** (small rename vs. OAuth refactor) — an application of the delegation gate, not data.

* **The general lever list** for token reduction, aside from the cited caching figure — real, common patterns shown without invented numbers.

## About this project

This guide is built with Mintlify. Its content was migrated from the original question-and-answer walkthrough. Published figures and their caveats are retained; diagrams are rendered with Mermaid.

[Back to the walkthrough](/learning-path)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.