---
source: https://iinventedlight.com/library/notes/ao-cost-compression-accounting.md
title: "Where context compression actually saves agent tokens"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2609.22114] An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents, CC BY 4.0; arXiv, [2609.26779] CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was rerun; only the fetched arXiv abstract pages were read. Results are as reported by the authors."
category: agent-operations
topic_area: "agent operations: cost"
tags: agents, cost, tokens, compression
sources:
  - id: S1
    title: "[2609.22114] An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents"
    publisher: arXiv
    url: https://arxiv.org/abs/2609.22114
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2609.26779] CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents"
    publisher: arXiv
    url: https://arxiv.org/abs/2609.26779
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Where context compression actually saves agent tokens

## Short answer
Measure each lever separately. In one gateway study, trimming tool schemas was the only reliably positive saving per turn; compressing file reads paid off over many turns. Another method cut cost up to 50% by only truncating or dropping content, never rewriting it. [S1][S2]

## Tips
- Filter unused tool definitions first. [S1]
- Do not summarize a summary; compact from originals. [S2]

## Claims
- id: C1
  claim: Tool-schema filtering was the only reproducibly positive lever.
  source: S1
  evidence: "the only unambiguously and reproducibly positive lever"
- id: C2
  claim: Compression savings grow over turns.
  source: S1
  evidence: "its cumulative saving grows quadratically"
- id: C3
  claim: CliffCompaction cuts cost up to 50%.
  source: S2
  evidence: "reduces cost by up to 50% under a bounded context"
- id: C4
  claim: CliffCompaction never rewrites content.
  source: S2
  evidence: "only truncating or dropping content, never rephrasing or rewriting it"

## Sources
- [S1] [2609.22114] An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents. arXiv. https://arxiv.org/abs/2609.22114 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2609.26779] CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents. arXiv. https://arxiv.org/abs/2609.26779 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).