Agent Cost Levers: cut cost per solved task without cutting quality

Review this session transcript [path]. List every large block of output (over 100 lines) that stayed in context after it stopped being useful, with the turn it appeared and the last turn it was needed.

Most long-session overspend shows up here first.

Here are the baseline and the new run for the same ten tasks [paste both]. Compute cost per solved task and the pass count for each, and the change.
If the pass count dropped, say so first. Do not call anything a saving if quality fell.

Run it after every lever you try.

# Agent Cost Levers: cut cost per solved task without cutting quality

A PaperChaseWebb pack, version 1.0. Written for Claude Code, Codex and Cursor, and plain Markdown, so the measuring sheet, the rules block and the prompts work with any agent harness that logs its usage.

Most advice on agent costs starts with the model price. That is the one number you control least. What you do control is how the agent spends its turns: how hard it thinks on easy steps, how much stale output it carries, how often it asks the model for something a command could answer, what it recomputes, and whether it knows when to stop. This pack turns those five habits into a measured routine. You record a baseline, pull one lever at a time, and keep only the changes that lower cost per solved task while the pass rate holds.

It comes from our own usage audits plus one public benchmark write-up we checked against its primary source and rebuilt in our own words. The source is credited by link in the Sources section.

## What is inside

| Section | What it is for |
|---|---|
| 1. The number to watch | Why cost per solved task beats tokens per call, and how to set up a ten-task baseline |
| 2. Lever one: thinking depth | Match reasoning effort to the step instead of running every step at maximum |
| 3. Lever two: context curation | Keep the working context small and current as a session grows |
| 4. Lever three: direct tool routing | Send deterministic questions to commands, not to the model |
| 5. Lever four: cache and reuse | Keep a stable prompt prefix and stop paying twice for the same work |
| 6. Lever five: verified stop | Define done as a passing check, so the agent stops polishing |
| 7. Claim discipline | How to report a savings number someone else can check |
| 8. Rules block | A paragraph for CLAUDE.md, AGENTS.md or a Cursor rule |
| 9. Prompts | 10 prompts to measure, diagnose and apply the levers |
| 10. Sources | What we used and how we used it |