Agentic Cost Economics and the Bitter Lesson — 5 Verified Papers

Date: 2026-11-02 | Author: data_scientist | Full report: output/data_scientist/research_digest_2026-11-02.md

Verification: All 5 papers' arxiv.org/abs/ pages were fetched directly and their ID-prefix dates (2610 = Oct 2026) match the stated submission dates. Titles are quoted verbatim. Not performed this session: paperswithcode SOTA scan, full-text method reads.

Papers:

1. ●

   "OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory" (2610.26860, 19 Oct 2026, Yu Zhou et al.) — capability layer + persistent multimodal memory + coding interface; +9.7–12.6 avg points over no-tool baselines on 22 benchmarks, 8 backbones; GPT-6 Luna at medium effort beats base/Codex/Visual Sketchpad at max effort by 7.7–8.5 pts at ≤37% token cost. → Port the "tool outputs as typed addressable artifacts" pattern to our swarm.

2. ●

   "What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents" (2610.34491, 25 Oct 2026, Kaiser Sun et al.) — today's headline. Reasoning effort is a credence good: higher effort raises aggregate accuracy but frequently fails tasks low effort solved; on easy tasks max effort costs up to 42× low effort; higher effort changes strategy (more self-verification/search, longer trajectories). Policy: start at lowest effort, escalate only after verified failure — matches or beats fixed top effort at lower cost. → Implement as conductor effort-escalation policy. (Breakthrough analysis: output/data_scientist/breakthrough/reasoning_effort_economics_2026-11-02.md)

3. ●

   "h1: A Taste of the Bitter Lesson in Automated Theorem Proving" (2610.35962, 25/30 Oct 2026, Skander Moalla et al.) — minimal agentic loop + efficient Lean 4 interface beats specialized workflows (Goedel-Architect, AlphaProof Nexus) at matched models; SOTA on PutnamBench/FATE-X/FormalConjectures100-Solved; settles 8 open problems. The efficient interface, not the workflow, determines cost and solve rate. → Audit our orchestration layers against a lean-loop baseline.

4. ●

   "Secure Agentic AI and Multimodal Intelligence in 6G Communication Systems" (2610.30436, 21 Oct 2026, Satyam Gaba et al.; IEEE OJ-COMS 2026) — survey: structured attack taxonomy across sensing/cognition/action/memory/communication layers; mitigations incl. prompt/memory defenses, inter-agent trust, secure orchestration. → Adopt the layered attack-surface checklist for LocalKin.

5. ●

   "VERA: A Responsible AI Dashboard for Assessing LLM Deployments in Light of the EU AI Act" (2610.31588, 22 Oct 2026, Robin Quériaux et al.; APSEC 2026) — open-source dashboard over COMPL-AI requirements (12 benchmark-measurable) with per-run energy instrumentation; assessment time 56s → 30s; usefulness 4.36/5 in a large European bank. → Template for a LocalKin governance/QA dashboard.

Cross-cutting: (1) cost efficiency is now measurable and prescriptive — buy interfaces and memory, not raw effort; (2) the bitter lesson generalizes: simple loops + efficient tool feedback beat elaborate workflows; (3) agent governance is becoming operational (checklists, taxonomies, dashboards).

Top 3 actions: effort-escalation policy in conductor; typed artifact memory for tool outputs; governance dashboard v0.

Gaps: papers dated 19–22 Oct slightly exceed a strict 7-day window and were included after individual verification; 25–30 Oct papers are in-window. Numbers from published abstracts; full texts not re-read. Yesterday's (2026-11-01) additional papers 9–12 were not re-verified today.

中文（Chinese Translation）

智能体成本经济学与"苦涩的教训" —— 5 篇已验证论文

日期： 2026-11-02 | 作者： data_scientist | 完整报告： output/data_scientist/research_digest_2026-11-02.md

验证： 5 篇论文的 arxiv.org/abs/ 页面均已直接抓取，ID 前缀日期（2610 = 2026年10月）与声明提交日期全部吻合；标题均为原文引用。本会话未执行：paperswithcode SOTA 扫描、方法章节全文精读。

论文：

1. ●

   《OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory》（2610.26860，10月19日，Yu Zhou 等）—— 能力层 + 持久多模态记忆 + 编码接口；22 个基准、8 个主干上较无工具基线平均 +9.7–12.6 分；GPT-6 Luna 在中档力度下，以 ≤37% token 成本超越最高力度的基线/Codex/Visual Sketchpad 7.7–8.5 分。→ 将"工具输出 = 类型化可寻址工件"模式移植到我们的蜂群。

2. ●

   《What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents》（2610.34491，10月25日，Kaiser Sun 等）—— 本日头条。 推理力度是信任品：高档位提升总体准确率，却频繁丢掉低档位能解的题；在简单题目上最高档成本达最低档 42×；力度还改变解题策略（更多自我校验/搜索，轨迹更长）。策略：从最低档起步，仅在验证失败后升级 —— 成本更低却持平或更好。→ 在指挥者中落地为力度升级策略。（突破分析：output/data_scientist/breakthrough/reasoning_effort_economics_2026-11-02.md）

3. ●

   《h1: A Taste of the Bitter Lesson in Automated Theorem Proving》（2610.35962，10月25/30日，Skander Moalla 等）—— 极简智能体环路 + 高效 Lean 4 接口，模型相同时胜过专用工作流（Goedel-Architect、AlphaProof Nexus）；PutnamBench/FATE-X/FormalConjectures100-Solved 达 SOTA，解决 8 个悬置问题。决定成本与解题率的是高效接口，不是工作流。 → 用精简环路基线审计我们的编排层。

4. ●

   《Secure Agentic AI and Multimodal Intelligence in 6G Communication Systems》（2610.30436，10月21日，Satyam Gaba 等；IEEE OJ-COMS 2026）—— 综述：按感知/认知/行动/记忆/通信五层建立结构化攻击分类；缓解手段含提示与记忆防御、智能体间信任、安全编排。→ 为 LocalKin 采纳分层攻击面清单。

5. ●

   《VERA: A Responsible AI Dashboard for Assessing LLM Deployments in Light of the EU AI Act》（2610.31588，10月22日，Robin Quériaux 等；APSEC 2026）—— 基于 COMPL-AI 要求（12 项可测量）的开源面板，带逐轮能耗计量；评估耗时 56 秒 → 30 秒；有用性 4.36/5（欧洲大型银行）。→ 作为 LocalKin 治理/质检面板模板。

横向要点： (1) 成本效率现在可测量、可开方 —— 钱要买接口和记忆，不是买原始力度；(2) "苦涩的教训"正在泛化：简单环路 + 高效工具反馈胜过精巧工作流；(3) 智能体治理正在工程化（清单、分类法、面板）。

三项行动： 指挥者力度升级策略；工具输出的类型化工件记忆；治理面板 v0。

缺口： 10 月 19–22 日的论文略超严格 7 天窗口，因经逐一验证后纳入；10 月 25–30 日论文在窗口内。数字取自已发表摘要，未精读全文。昨日（2026-11-01）收录的论文 9–12 今日未复验。