Context is the cost: Stateless Language Agents, SelfCompact, Fork-and-Flush, ReHarness

Sources: HuggingFace Daily Papers 2026-10-12, Apple ML Research (lab watcher), The Decoder (RSS), X Following feed. Raw: raw/huggingface/2026-10-12-stateless-language-agents-scaling-long-horizon-automated-res.md, raw/huggingface/2026-10-12-recurrent-harness-for-agentic-continual-learning.md, raw/labs/2026-10-13-103914-labs.md, raw/rss/2026-10-12-the-decoder-clearing-chat-history-helps-ai-agents-find-better-solut.md.

TL;DR

Four results on the same question: what should a long-running agent keep in its context window?

• Stateless Language Agents (SLA) (arXiv 2610.07625, Stanford, CMU, UW, SambaNova, Berkeley). "Stateful search with stateless agents": no agent keeps its conversation between calls. The harness owns the research state (candidate solutions and measured scores) and builds a fresh, role-specific context for each call. A stateless Advisor reads harness-summarized evidence and assigns concrete experiments to parallel Workers. At budgets up to 1 billion tokens on software engineering, kernel optimization and algorithm design, SLA gets the best final result on every task and matches the strongest kernel baseline's final score with 84% fewer tokens; the Advisor uses under 0.6% of tokens. Also argues short-budget evaluations misjudge research systems because benchmarks saturate early.
• SelfCompact (Apple ML Research, NeurIPS, with JHU). Let the model decide when to compact: a summarize tool plus a short rubric saying when to fire (a sub-task resolved, the trajectory converging) and when not to (mid-derivation, stuck). Neither works alone. Up to +18.1 points on math and +5-9 on agentic search over no summarization, at 30-70% lower per-question cost, matching fixed-interval compaction at a fraction of the tokens. No fine-tuning.
• Fork-and-Flush (Microsoft and Dimitris Papailiopoulos; The Decoder, thread). Long-running research agents fall into "idea basins": they keep polishing one family of ideas because their growing context anchors them. The fix: periodically fork into copies that share the workspace but start with empty chat history, then continue from the best. Up to 66% better results at equal compute, beating single runs and best-of-N on 12 of 13 long-horizon tasks.
• ReHarness (arXiv 2610.12668). For continual learning over a stream, revise past reflections using each new trajectory (backward flow through the textual harness state), not only write new ones. Beats raw-context, memory, reflection and prompt-optimization baselines on three benchmarks and three backbones.

flowchart LR
  H["Harness state<br/><small>solutions + scores</small>"] --> A["Advisor<br/><small>fresh context, assigns</small>"]
  A --> W["Workers<br/><small>stateless, parallel</small>"]
  W -->|results| H
  W -->|context rotting| C["Compact or flush<br/><small>rubric or fork</small>"]
  C --> W
  H --> O["Best result<br/><small>84% fewer tokens</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class H input
  class A,W core
  class C loop
  class O exit
  linkStyle 2 stroke:#f08c00,stroke-width:2px

Also in this cluster

• Persistent memory null result (arXiv 2610.07782, UCLA et al., DAIR pick). A three-tier chain-of-agents cuts peak KV cache to 14.3 MiB per query (vs ~35 MiB single-pass/RAG) and latency 52-67 s to 29 s. The persistent tier that stores reasoning traces adds +0.368 MiB and +0.015 accuracy with a CI crossing zero. Decomposition pays; the memory tier does not on single-question benchmarks.
• DUET (DAIR, arXiv 2610.04087, Microsoft). Update the LLM grader and the agent in alternating rounds. Finance Agent Benchmark 96.5% vs GEPA 94.4%; grader agreement on GDPval 68.4% to 78.5%. Training the grader with the solver frozen barely moves agreement. Same idea as Double Ratchet (10-13 Media Zone): the grader must evolve with the agent.
• Agent Arena "Confirmed Success" (post). July's leader (Claude Fable 5) is now the top-20 average after 12 weeks; Kimi K3 matched it in 3 weeks at 89% lower cost per task, DeepSeek V4.1 Flash at 96% lower. Top 5 now confirm 94% of tasks done, up from 80%.
• Continuous agent scoring (Ken Huang, Agentic AI). Android Bench 2.0's long-horizon tab: Claude Opus 5.5 on Claude Code at 32.7% pass but 84.7% completion, average 15.2 hours and $215 per task. Binary pass/fail erases 90%-done work; report a continuous completion score alongside it.

Relation to prior wiki pages

• Pattern (n=4 this week): context is a liability past a point. GitSwarm (10-12) put durable state in a shared Git repo; REMORY (10-10) compacted context into learned soft tokens; today SLA and Fork-and-Flush reset context entirely while keeping the workspace, and SelfCompact lets the model choose when. The common rule: state belongs in the harness or the workspace, not in the transcript.
• Cost link: SLA's 84% token saving and SelfCompact's 30-70% cost cut are the agent-level version of the kv-cache page's session thread. Shorter contexts are fewer KV bytes per step.
• Anthropic's self-replication post (10-13) uses the same harness trick (Memento restarts fresh agents on the previous workspace). The method now appears in research automation, compaction and red-teaming.

Related

agent-harness-engineering · agent-memory · self-evolving-agents · agent-benchmarks