SparseRead cuts agent reading tokens by up to 92.9% without quality loss
ict-cas · hf · 2026-11-12
ict-cas released SparseRead, a model-agnostic reading control layer that stops web-facing agents from ingesting entire pages, docs, repos, and tool outputs when only sparse evidence is needed:
- Read Intent Protocol tracks evolving information needs and source coverage
- Adaptive Reader Backends decide how evidence is located and exposed
- Budgeted Read Gate bounds cumulative raw exposure per episode with faithful full-view fallback
Tested across 6 models (including Claude Opus 5) and 5 scenarios: tokens drop up to 92.9% (median 69.0%), wall time up to 89.0%, and quality holds or improves in 97% of settings. Web deep-search tasks see 31–43% token savings and 23–37% faster completion. Integration with 3 agent frameworks confirms portability. Open-sourced on GitHub/HF.
More from coding & agent
- TerraPulse: a global LiDAR archaeology platform inspired by 48-agent jungle scan of 1,200 km² — RichardsonDx · 2026-11-12
- Mashvisor ships an MCP server for rental property analytics, with setup walkthrough — ahmadhashlamoun · 2026-11-12
- Andrew Ng releases 4 free AI tools on GitHub, including desktop coworker OpenWorker — alex_verem · 2026-11-12
- The fastest way to make $0 with AI: automate outreach before your first client — aryanXmahajan · 2026-11-12
- Slaunt MCP server gives Claude Code 99% memory recall while cutting tokens 31% — karatyma_dev · 2026-11-12
- Managed agents platforms are the new chat-with-a-PDF, says developer — intellectronica · 2026-11-12