ParSer from CUHK decouples document reading from question reasoning in long-context QA. Instead of the MemAgent-style loop where each chunk update depends on the previous one, they pin one frozen Qwen3.5-4B subagent per chunk and have them all answer the lead agent's query in parallel; the lead agent — blind to the document — runs a ReAct loop, scatter-gathers findings, and refines queries per round. Only the lead agent is trained, via GRPO with exact-match reward. Cross-round KV-cache reuse keeps the re-reading cheap. The results are the interesting part. A 4B lead beats the strongest sequential memory baseline by 5.7 points average and 12.0 at 896K tokens; a 9B lead beats DeepSeek-V4-Pro's full 1M-token context by 6.3. The controlled perturbation experiments are the cleanest evidence: sequential memory agents swing badly when you move evidence position, reverse evidence order, or spread evidence apart; ParSer stays nearly flat. Latency drops 11× at 896K (78s vs MemAgent's 876s), still 1.7× ahead at concurrency 16. Caveats worth noting: lead agent never sees the document, so its RL policy generalizes but the chunk readers do the heavy lifting; and the latency numbers come from a two-GPU ParSer deployment versus one-GPU baselines, which they at least flag honestly. Training used 16 H100s total, for scale planning. The obvious open question: how does the query-formulation policy hold up when evidence isn't decomposable into chunk-local hops — multi-hop questions whose hops span chunks semantically rather than positionally?