DTOC paper: agents hide tool outputs to cut tokens and boost solve rate
pppeer · reddit · 2026-10-03
A new paper (Discovery Science 2026, authors incl. Abhay Chaturvedi and Peter van der Putten) introduces DTOC: Dynamic Tool Output Compression, a simple context-management scheme for agents.
- In every turn the agent can hide or unhide specific tool outputs to focus context; outputs are persisted separately so they can be re-enabled later—no irreversible lossy compression
- On a sample of DeepSWE tasks plus internal ablations, DTOC saved tokens, steps, and cost while improving solve rate, though results are model- and task-dependent
- A reference implementation in OpenCode is available; the authors invite practitioners with similar approaches to share
Related event: DTOC Paper: Dynamically Hiding Tool Output Saves Tokens and Boosts Accuracy(2 posts)→
More from coding & agent
- GPT learns to pass tests, not to engineer: the RL reward-mismatch behind ugly code — xiaohu · 2026-10-03
- Dev wires a daily Hermes agent to triage Sentry errors, but it keeps 'forgetting' context — pauljasperdev · 2026-10-03
- Runway MCP one-shots a corporate training video in 15 minutes, typically $10k to make — tlakomy · 2026-10-03
- Building an AI Research System for a One-Person Firm: Progressive Determinism Architecture — otionflow · 2026-10-03
- Dev deletes custom CLI after discovering SmartSub v3.9.0 already shipped CLI + MCP — lxfater · 2026-10-03
- Meta reportedly building Agent Engine codenamed Forge for its API platform — testingcatalog · 2026-10-03