SparseRead: read less, solve more
TL;DR
Agents read whole web pages, files, repos and tool outputs when they need a few lines. That over-reading costs tokens and latency, and it dilutes the evidence the model needs. SparseRead is a layer that sits between the agent and its sources and decides what text is admitted, without touching the model. It has three parts: a Read Intent Protocol that tracks what the agent currently needs and what each source has already covered; Adaptive Reader Backends that pick how to locate and expose evidence (search, outline, slice); and a Budgeted Read Gate that caps the raw text exposed per episode, with a full-view fallback when the sparse view is not enough. Across six models (including Claude Opus 5) and five scenarios it cuts total tokens up to 92.9% and wall time up to 89.0%, preserving or improving quality in 97% of settings. On web deep-search it saves 31 to 43% of tokens and 23 to 37% of time.
flowchart LR
A["Agent<br/><small>any model</small>"] --> I["Read intent<br/><small>need and coverage</small>"]
I --> B["Reader backends<br/><small>search, outline, slice</small>"]
B --> G{"Budget gate<br/><small>cap per episode</small>"}
G -->|sparse view| A
G -->|not enough| F["Full view<br/><small>faithful fallback</small>"]
F --> A
X["Sources<br/><small>web, docs, repos, tools</small>"] --> B
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,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
class X input
class A core
class I,B,G loop
class F exit
linkStyle 3 stroke:#2f9e44,stroke-width:2px
Key findings
- Token savings come from admission control, not compression: the agent sees less raw text, not a summary of it.
- The fallback keeps it faithful: when the sparse view fails, the full object is still reachable.
- Portable across three agent frameworks; no model changes.
Gaps
- Savings are largest on scenarios with big objects; on deep search they fall to 31 to 43%.
- No breakdown of how much comes from prompt-cache interplay: a sparse view may break prefix caching that a stable full view would hit.
How it relates to prior wiki pages
- Extends agent harness engineering: the harness, not the model, controls context cost. It is the read-side cousin of context compaction (Microsoft's CompactionBench, mentioned the same day) and of AWS's Bedrock cost-per-task guidance naming agent-loop turns and verbosity as hidden cost drivers.
- Connects to the token-spend data in agent memory and the 11-12 WikiTrace page: CMU/JetBrains found doubling token spend buys about 10% more commits. Read control is one of the few levers that cuts spend without cutting output.