arXiv:2610.31710v2 Announce Type: replace-cross Abstract: Context compaction lets long-horizon coding agents continue beyond their context windows, but can erase the state needed to finish a task. Prior evaluations measure efficiency, retained information, or end-to-end success, leaving the effect of a specific history replacement difficult to isolate.

Recommended

Prompts to Proxies: Emulating Human Preferences via a Compact LLM EnsemblePrompts to Proxies: Emulating Human Preferences via a Compact LLM EnsemblearXiv cs.AI Evaluating and improving AI agents for sequencing protocol understandingEvaluating and improving AI agents for sequencing protocol understandingarXiv cs.AI Answer-Only Disclosure: Private-Context LLM Assistance over Web APIs with a Leakage LedgerAnswer-Only Disclosure: Private-Context LLM Assistance over Web APIs with a Leakage LedgerarXiv cs.AI Encode Once, Answer Many: Internalizing Context via Parametric FlowEncode Once, Answer Many: Internalizing Context via Parametric FlowarXiv cs.AI Where You Inject Matters: The Trust Region Effect in LLM AgentsWhere You Inject Matters: The Trust Region Effect in LLM AgentsarXiv cs.AI WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form VideoWeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form VideoarXiv cs.AI 10 best AI observability tools for monitoring and evaluating agents in 202610 best AI observability tools for monitoring and evaluating agents in 2026Mintlify Pancakes Are Delicious and Data Centers Are for Free StuffPancakes Are Delicious and Data Centers Are for Free StuffThenewstack How We Eliminated Long-Lived CI Secrets Across 70+ ReposHow We Eliminated Long-Lived CI Secrets Across 70+ ReposPulumi Pulumi Neo Now Supports AGENTS.mdPulumi Neo Now Supports AGENTS.mdPulumi Go Concurrency Patterns: Context - The Go Programming LanguageGo Concurrency Patterns: Context - The Go Programming LanguageGo How a top tier European soccer team sped up their data processing and reduced costs by 50%How a top tier European soccer team sped up their data processing and reduced costs by 50%Modal How to serve trillions of tokens for trillion-parameter coding agentsHow to serve trillions of tokens for trillion-parameter coding agentsModal Do AI Agents Need a Semantic Layer?Do AI Agents Need a Semantic Layer?Motherduck Building a Remote MCP Server: OAuth, Tool Design & Lessons from 4,000+ AI Queries | MotherDuckBuilding a Remote MCP Server: OAuth, Tool Design & Lessons from 4,000+ AI Queries | MotherDuckMotherduck SQL is Dead, Long Live SQL: Engineering a Reliable Analytics Agent from ScratchSQL is Dead, Long Live SQL: Engineering a Reliable Analytics Agent from ScratchMotherduck Cerebras Systems Enables GPU-Impossible™ Long Sequence Lengths Improving Accuracy in NaturalCerebras Systems Enables GPU-Impossible™ Long Sequence Lengths Improving Accuracy in NaturalCerebras The Livecycle Docker Extension: Instantly Share Changes and Get Feedback in ContextThe Livecycle Docker Extension: Instantly Share Changes and Get Feedback in ContextDocker AI's Role in Video Marketing, Cost EfficiencyAI's Role in Video Marketing, Cost EfficiencyHeygen AGENTS.md outperforms skills in our agent evalsAGENTS.md outperforms skills in our agent evalsVercel