Top developments
The agent-harness cost thesis sharpens across three papers in this week's NLP Newsletter roundup. OO-Harness evolves specialized harnesses as subclasses and routes to them, lifting AutomationBench from 61.6% to 77.4% with the same Gemini 3.5 Flash executor; Cache Engineering (AKV) lets the agent control its own KV cache, cutting cost 27.7% versus summarization; and JevAgentHarness delegates routine tool choices to a fitted 4B decision model at 81% to 89% lower cost. The lever for cheaper long-running agents is increasingly the harness and serving layer, not the next model checkpoint. Genuinely useful, with the usual single-benchmark caveats, and note that no public paper link for OO-Harness itself surfaced yet: it is reported here via the newsletter roundup and the Knowledge Vault entry filed today.
MemoryTax: stop overengineering agent memory (Shanghai Jiao Tong). Across seven memory approaches on one 419-turn LoCoMo conversation, Mem0 only won with the weakest model (GPT-4o-mini); with Claude Opus 5.5, agent-directed search of raw history hit 99% for under half the full-context cost, and memory quality was capped by the model that wrote it (author page). Start with full context, add a memory layer only where reuse justifies it, and keep original records. Caveat: one conversation, one run per config, so treat it as a framework, not a verdict.
OpenAI is reportedly discounting enterprise deals 20% to 40% to win Anthropic customers. Reporting (The Information, via SiliconReport) names Rippling (OpenAI coding share 16% to 52%, Anthropic 57% to 18%), Replit (30% or more off a contract worth over $100M a year), and DataCamp (claimed 80% to 90% cost reduction). Model choice for many everyday workloads is becoming a procurement negotiation, not a capability decision. Commercial signal, not a technical one, and the savings claims are vendor and customer statements without independent audit.
Frontier AI labs
No new frontier model releases in the window. The Oct 28-31 tracker pegs mentioned yesterday remain unconfirmed projections. Kimi K3.1 Pro weights are still unpublished ahead of the promised Nov 4 date; that remains the watch item, not a done deal.
Four Gemini 4 Argon agents beating Minecraft (DeepMind) headlines the week's paper roundup. Four agents, no assigned roles or human-written game logic, defeated the Ender Dragon in 27 hours and 14 minutes after writing 2,783 programs and dying 605 times (Tom's Hardware). A long-horizon coordination datapoint, not a game stunt; the caveats (single run, Easy difficulty, Mineflayer exposing game state) are stated by the authors. Already in the Knowledge Vault from earlier coverage.
AV / embodied-AI research
Quiet day on AV. No new AV papers in the window, and neither Bluesky account posted anything new in the window (latest jlake9 post is Oct 14; latest ericzzj posts are Oct 23, already covered). Wayve/Stellantis/NVIDIA France and Rivian/Third Dimension AI were covered Oct 24 and are not repeated as news.
Closest embodied item: Agent S4 (Simular) lets a computer-use agent build its own interfaces from task experience. A meta-agent evolves interfaces plus usage notes without success labels; with Qwen3.8-27B it reaches 21.3% on OSWorld V2 (3.8x fixed interfaces, above GPT-5.5 and Claude Opus 4.8 at a fraction of their cost) and lifts ScienceBoard from 41.4% to 58.6% (arXiv). Interface design as an agent capability is a genuinely new improvement axis, though OSWorld scores remain low in absolute terms.
Research/papers worth reading
OO-Harness (Google Cloud AI Research / UIUC). Evolved harness subclasses plus a test-time router; inheritance itself added about 5 points on Workspace-Bench versus flat evolved code. No public paper link surfaced for this item yet, so it is left unlinked here and reported via the newsletter roundup.
Cache Engineering for Agents / AKV (Tsinghua / CUHK). Drop and reposition KV-cache operations with a stateless serving interface; up to 1.2x throughput, 1.5x faster P95 latency (MarkTechPost). Filed to the vault.
Self-evolving continual learners (Google DeepMind / Mila). Across 12 method/scenario pairs, plasticity and stability correlate at r = -0.72: self-evolving agents improve on streams but forget, via global prompt edits and cross-domain memory collisions (DAIR.AI). Worth reading if you build self-improving agents; static benchmarks hide both failure modes.
DiscoveryRSI (PhAI Labs / CUHK / Stanford / Harvard Medical School). Recursive self-improvement for scientific hypothesis generation; automated hypothesis quality rose from 4.86 to 5.39 after two rounds, with DFM-Bench released (arXiv). Interesting for research automation, less directly relevant to AV.
Transcoder features are neurons (Anthropic / Georgia Tech). Many late-layer transcoder features are close copies of single MLP neurons, casting doubt on non-neuron features as interpretability evidence (arXiv). Niche but a useful corrective.
Agents/tools/models
JevAgentHarness / AgentDecision-4B (UIUC / Princeton / Google). Fitted confidence thresholds let a decision model take routine choices with no outcome change on 500 tasks (14.2% lower cost); the open 4B model trained on post-mistake states beats Jev itself on out-of-distribution tasks (91.5% versus 82.0% solved) (arXiv). Filed to the vault.
Terminal-RSI (Tencent). Alternating fine-tuning and PPO from Qwen3.8-27B lifts Terminal-Bench average from 20.68% to 50.58%; a reward audit found dense progress rewards gave failed runs 50% of reward, cut to 28.7% by separating strict success from partial credit (arXiv). Useful methods note on reward design.
Indra OCR 1.1 (Datalab). A 1M-parameter OCR model claiming 44 pages per second on a 32-core CPU, 15x faster than Tesseract, 80+ languages, layout plus line boxes (Datalab). Small and CPU-friendly is the story; note this arrived in a sponsored slot, so treat the benchmarks as vendor claims until independently checked.