What the agent sees decides what it costs: TokenEconomizer, CQ2-CARE, CarryOn, Librarian Agents

Sources: HuggingFace daily papers 2026-10-19: TokenEconomizer, 2610.22732; Skip the Compressor (CQ2-CARE), 2610.21073; Pack Light, CarryOn, 2610.18784; Librarian Agents, 2610.22273. Raw: raw/huggingface/2026-10-19-* (see each slug).

TL;DR

Four papers on one question: which tokens should an agent's context hold?

• TokenEconomizer runs coding agents over sequences of real tasks in the same repo and finds recurring, environment-specific token waste. An external controller decides which micro-skills to inject, which persistently unused tools to prune, and when to break unproductive failure loops. 18.4-30.9% fewer tokens with comparable success across Claude Code, Codex and Copilot CLI; 6.4-21.2% on web dev and security audits.
• Skip the Compressor (CQ2-CARE): at an equal token budget, picking whole files verbatim beats prompt compressors by 6.3 points and retrievers by 9.3-12.7 points on 300 SWE-rebench issues (43.7% with GPT-5.6, 46.7% with Opus 5). Compressors keep the right files but alter the text, so patches fail to apply.
• CarryOn organizes long-horizon context around reusable task objects ("carriers") instead of history, updating them at each fold. +5.8% on OSWorld 2.0 with 51.9% fewer tokens; +7.0% on BrowseComp-Plus with 22.9% fewer.
• Librarian Agents trains a builder agent with RL to turn a raw corpus into a filesystem workspace that a later answering agent uses. +33 points on BrowseComp-Plus over the best retrieval baseline; +31 points zero-shot on BRIGHT; build cost amortizes over many queries.

flowchart LR
  C["Raw context<br/><small>repo, history, corpus</small>"] --> X["Compressor<br/><small>deletes text</small>"]
  C --> S["Selector<br/><small>whole files, objects</small>"]
  X --> F["Patch fails<br/><small>lines altered</small>"]
  S --> K["Agent context<br/><small>verbatim, smaller</small>"]
  K --> O["More solved<br/><small>18-52% fewer tokens</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  class C input
  class X err
  class S loop
  class K core
  class F err
  class O exit
  linkStyle 2 stroke:#e03131,stroke-width:2px
  linkStyle 4 stroke:#2f9e44,stroke-width:2px

How this relates to prior wiki pages

• Pattern now n=5 on "selection beats compression": Memorilla (10-19, 16 memory tokens beat full context on FactKG), Meta's Thin Orchestrator (today, isolate each specialist's context), CQ2-CARE, CarryOn and Librarian Agents. The common claim: deciding what enters context dominates shrinking what is already there.
• Tension with Memorilla: Memorilla compresses to soft tokens and wins on fact verification; CQ2-CARE shows compression loses when the task needs exact text. The 10-19 Looking Ahead predicted Memorilla would break on exact recall; CQ2-CARE is indirect evidence for that prediction, on code.
• TokenEconomizer extends the 10-19 token audit of security agents (which found 60.4% of tokens go to understanding code) from measurement to control. See agent-harness-engineering and agent-memory.

Gaps

• TokenEconomizer and CarryOn report averages; variance across repos and the cost of the controller itself are not in the abstracts.
• Librarian's upfront build cost vs query volume break-even is not quantified in the abstract.