推理成本、延遲最佳化、可觀測性、企業落地、規模化
系統性測試 5 個 frontier 模型是否具備「預算感知」能力，結果全員失敗：任務能力強不等於預算感知強（r=0.35），所有模型都過度樂觀，在注定失敗的任務上持續燒錢。
arxiv: 2606.00198把 agent 安全問題拆成四個細粒度維度打分，用這個「評分標準（Rubric）」作為強化學習的獎勵訊號，讓模型學會在安全和實用之間找到真正的平衡，而非一律拒絕。
arxiv: 2606.04051Deep research agent 在評測時會搜尋網路，而 benchmark 答案也在網路上——這種「邊考試邊查答案」現象讓分數最多虛高 4%；Alibaba 團隊定義三種汙染類型並開發偵測演算法，質疑現有評測數字的可信度。
arxiv: 2606.05241一個 AI「優化 agent」不只幫你寫更好的 system prompt，也同時把自己的 system prompt 一起優化——不改模型，在五個 benchmark 上平均提升 4.49 分。
arxiv: 2606.04465現有 LLM serving 模擬器把每個對話 turn 當獨立請求，完全忽略 agent 跨輪次的 KV cache 重用和工具呼叫等待；這篇建了第一個把 agent 當「有狀態程式」來模擬的 hardware-aware 模擬器。
arxiv: 2606.09613訓練 Agent 時大部分的「跑樣本」都在浪費算力；TRACE 用樹狀結構把 rollout budget 導到最有學習價值的岔路口，同樣算力讓 Multi-Hop QA 多學 2.8 分。
arxiv: 2606.11119agent 系統還不穩定就上線時，傳統任務層監控會失靈；這篇提出 3×3 監控框架，幫你找出「哪根線沒接好」。
arxiv: 2606.02494LLM Agent 明明答對了卻說「我答錯了」——本文用一個不需額外標注的校準獎勵修正這個反思偏差，讓 Agent 真正成為自己的驗證器。
arxiv: 2606.14211小模型面對 250 個工具的 MCP 目錄幾乎癱瘓（執行成功率 3%），Evoflux 用推論時的演化搜尋反覆修復失敗的工具 graph，把成功率拉到 17–24%——不需要 fine-tuning。
arxiv: 2606.12674Agent 常常「以為自己知道」使用者的意思，其實只是靠假設在行動；ACCORD 讓 agent 在每個行動前主動去環境裡確認「我真的有這個資訊嗎？」，AppWorld 成功率從 42% 跳到 62.6%。
arxiv: 2606.16432Perplexity 的真實生產數據：把功能從「對話搜尋」升級為「自主 agent」之後，每次任務自主執行時間從 33 秒跳到 26 分鐘、完成時間縮短 87%、用戶滿意度更高——而且 agent 吸引了認知複雜度更高的工作類型。
arxiv: 2606.07489工具呼叫 agent 常常「知道事實卻用錯」：資訊都在 prompt 裡，但每次決策時 LLM 需要自己從文字中重建狀態，一旦重建出錯就違反政策或拿陳舊資料做決定。LedgerAgent 在 agent 迴圈中插入一個明確的型別帳本，把狀態從 prompt 文字裡分離出來，不需要任何微調。
arxiv: 2606.20529企業部署 agent 時，「批准一個動作」和「執行那個動作」之間存在安全漏洞：現有存取控制在 identity 層，assurance 在 plan 層，但在「agent 真的動 AWS / K8s 的那一刻」缺乏強制驗證點。Sovereign Execution Broker (SEB) 就是插在這個瞬間的憑證驗證與執行邊界。
arxiv: 2606.20520把 LLM multi-agent 系統推向企業生產的兩大卡點是「Domain 適配」和「推理成本」，這篇提出兩階段框架：先客製化（持續預訓練 + SFT + 偏好最佳化），再加速推理（推測解碼 + FP8 量化），實測吞吐量提升 4.48 倍。
arxiv: 2606.18502Agent 跑長任務時 trace 越積越長最終爆 context，SelfCompact 讓 LLM 自己判斷何時壓縮記憶，比固定間隔方案更準確，且節省 30-70% token 成本。
arxiv: 2606.23525Web Agent 執行完任務後，HANSEL 自動從瀏覽歷史中抽出「最關鍵的幾頁」讓用戶點進去驗證，比起讓人看完整 log 省了 61% 的軌跡量。
arxiv: 2606.18671精心訓練的 agent（不論 SFT 還是 RL）放到和訓練時不一樣的環境，表現就會明顯下降——這不是個別模型問題，而是系統性問題。
arxiv: 2607.01084不改模型、零訓練，推理時用「模型自己的注意力分數」遞迴挑出相關片段重播給模型，顯著改善 LLM 在 128K 長文中找不到答案的問題。
arxiv: 2607.02509問題不是 RL 演算法不夠好，而是缺乏讓 agent 在生產中「邊跑邊進化」的系統基礎設施；這篇提出三大工程支柱讓企業規模的 agent 能持續自我更新。
arxiv: 2607.01120Amazon 把 agent 每次都要「重新生成」的固定操作程式碼，預先離線編譯成版本化工具——生產環境 p50 延遲降 42%、錯誤率降 53%。
arxiv: 2607.08010另開一個「記憶 agent」並行跑在行動 agent 旁邊，在關鍵時刻主動把重要資訊推送給它——Terminal-Bench 2.0 提升 +8.3pp，τ²-Bench 提升 +6.8pp。
arxiv: 2607.08716爬梳 1,250 篇論文、建立「AI 自我改良」的雙軸分類地圖，分清楚哪些是已可上線的「有界自我精煉」、哪些是還在研究的「無界遞迴自我提升」。
arxiv: 2607.07663把 agent 評測拆成三個可獨立抽換的零件（題庫、執行框架、執行環境），解決現在大家各做各的、無法重用的問題，並附帶自動偵測 reward hacking（作弊行為）的軌跡分析工具。
arxiv: 2607.13705整合製藥與金融系統真實部署經驗的 agent 落地報告，歸納出三個讓 agent 從實驗室走向生產的關鍵設計模式：驗證管線、fallback 機制、人機協作節點（human-in-the-loop）。
arxiv: 2607.19336LLM Agent 在開放環境中會隨時間漂移（預測越來越不準）；這篇把 ARIMA 時序預測器嵌進 Agent 作為「自我校準器」，讓 Agent 在無人監督下自動偵測並修正漂移，準確率比基準 LLM Agent 高 91.7%。
arxiv: 2607.22400Callosum closes a $100M seed round led by Atomico, valuation undisclosed. The bet: the agent cost bottleneck is not the model itself but cramming every step into the same GPU. As inference spending eats over half of AI-native companies' revenue, the routing layer's value expands from model selection to chip selection.
Three papers probe the real-world limits of AI Agents from different angles: ORCA-bench drops LLM Agents into production SRE on-call for root cause analysis — the best model scores only 40%; AgentS4D reveals the safety blind spot of workspace agents — 66% of 'successful' runs still triggered dangerous behavior; a Context Files study finds that AGENTS.md / CLAUDE.md files show no measurable improvement in coding agent correctness across 288 controlled trials.
Three papers tackling core Agent challenges: TRACE-ROUTER shows per-call model routing breaks in multi-step agent flows and proposes task-level routing with RL; OmniaBench builds a 1,431-question benchmark spanning consumer, enterprise, and engineering scenarios where top models (Claude Sonnet-5) still score under 60%; a self-calibrating agent framework uses ARIMA time-series forecasting to detect and correct prediction drift without human supervision.
Three papers tackle core AI agent platform challenges from different angles: **AgentCompass** introduces composable open-source evaluation infrastructure to end the fragmentation of agent benchmarking; **Agents in the Wild** is a rare production deployment report distilling reusable design patterns from pharma and finance; **Nanbeige4.2-3B** proves a 3B model with Looped Transformers and large-scale agentic RL can outperform 9B and even 12B competitors on agent tasks — directly relevant for edge deployment and cost-sensitive scenarios.
Three papers approaching 'how to make agents reliably solve complex tasks' from complementary angles. NVIDIA proposes writing agents as plain Python classes so development, testing, and tracing work like normal software engineering. BAAI's AREX demonstrates a deep-research agent that recursively verifies and refines its own conclusions, outperforming comparable-scale models on BrowseComp, HLE, and other benchmarks. The third paper surveys 1,250 papers to build a clear taxonomy for the chaotic term 'AI self-improvement,' helping you tell which techniques are production-ready and which remain research-only.
Three papers tackle the same core question from infrastructure, observability, and evaluation angles: how do you build truly reliable agent systems? Dyserve uses mathematical optimization to decide which LLM each agent workflow node should use within 60ms, beating all baselines on both accuracy and latency. AgentLocate solves the ops nightmare of not knowing which agent broke a multi-agent pipeline, automatically pinpointing the responsible agent and the failure timestep (COLM 2026 accepted). PolyWorkBench delivers a warning: state-of-the-art LLM agents degrade significantly in multilingual workflows — global product scenarios still have a long way to go.
Three papers converge on one trend: the bottleneck for production agents is no longer model capability — it's state management. Paper 1 (Amazon) shows that pre-compiling repetitive steps into tools cuts p50 latency by 42% and error rate by 53%. Paper 2 introduces a standalone memory agent that proactively pushes critical state to the action agent, addressing behavioral state decay in long-horizon tasks. Paper 3 uses recursive multi-agent orchestration to overcome a single agent's inability to search both broadly and deeply. Together: **tool compilation, proactive memory, recursive orchestration** are the three pillars of agent platform engineering in 2026.
Three papers today reveal a core tension: current agent systems shine in closed environments but degrade sharply once conditions shift even slightly. An ICML 2026 paper systematically quantifies this problem through the lens of tool use; the second shows how a pipeline of 6 specialized agents can tackle complex cross-domain tasks; and the third reminds us from a UX perspective that agent 'personality intensity' isn't a case of more-is-better — moderate is the sweet spot.
Three papers tackling production-grade agent systems from different angles: a full-stack practical guide from LLM foundations to multi-agent architectures, a lightweight scaffold that lets agents decide when to compress their own context, and an RL training algorithm that refines credit assignment from tool-call boundaries down to the token level. Together they map out three key questions for building an agent platform: what architecture to learn, how to keep it stable at runtime, and how to train it better.
Three papers approaching agent reliability and safety in production from three layers: inference-time, training-time, and infrastructure. LedgerAgent uses a lightweight ledger structure at inference time so tool-calling agents no longer stuff all state into the prompt for the LLM to reconstruct — directly reducing policy violations and state errors. Alibaba's Connect the Dots (CoD) takes the longer view, using reinforcement learning to train agents that update their environmental awareness while executing tasks in long-term deployments, improving across tasks over time. Sovereign Execution Brokers tackle the security infrastructure layer, inserting credential verification at the exact moment an agent touches a production system, strictly binding authorized actions to actually executed actions. Three papers
Three papers challenging core assumptions about agent tool use and memory: Evoflux shows compact models nearly fail at MCP tool catalogs (3% success) and uses inference-time evolutionary search to reach 17-24%; FlowBank precomputes diverse workflow portfolios and routes at inference time, beating handcrafted designs by ~15%; GitOfThoughts reveals memory only helps when problems are near-duplicates (similarity > 0.8), but git version control offers an engineering path through auditability and replayability.
Three papers address agent reliability from three layers. RefGRPO fixes a neglected reflection calibration problem in agentic RL, turning agents into their own verifiers. 'Agents All the Way Down' delivers a complete custom-agent methodology from LLM substrate to production, arguing that solid foundations matter more than framework choice. EurekAgent uses autonomous scientific research to show that environment engineering beats process engineering for agent reliability.
Three papers tackling core Agent platform challenges from the angles of memory architecture, training efficiency, and reliability evaluation. HORMA proposes a hierarchical filesystem memory architecture so Agents stop collapsing under exploding context in long workflows; TRACE redesigns rollout budget allocation for Agent RL training, squeezing an extra 2.8 percentage points on Multi-Hop QA from the same compute; and τ-Rec exposes the 'reliability cliff' in multi-turn conversational recommendation Agents — even the strongest model drops to just 38% reliability over four consecutive runs, a sobering number for any team planning to ship an Agent product.
Three papers today approach agents from two angles — how to evaluate them and what they fundamentally are: T1-Bench introduces a high-fidelity benchmark spanning 25 real business domains, giving cross-domain reasoning its first systematic quantitative baseline; VISTA solves the credibility problem of using LLMs to simulate users for agent testing, providing 6 metrics to quantify whether your tests actually cover the agent's capability boundaries; Agentic Software clarifies from first principles that when the LLM becomes the primary reasoning engine, the nature of software has changed — directly impacting how agent platforms should design their debugging tools and testing strategies.
Three papers today explore 'agent-native infrastructure' at different layers: the first redesigns API error responses to give agents structured recovery hints, dramatically improving tool-call success rates; the second argues Agent OS is the right abstraction for long-running agents; the third builds a hardware-aware simulator for multi-turn agent serving to quantify KV cache scheduling trade-offs. From APIs to OS to hardware, every layer of the agent stack needs rethinking.
Three papers tackling core agent platform gaps from three angles: APB introduces a 4,209-question diagnostic benchmark that separates planning failures from execution failures; MetaForge lets agents forge missing tools at runtime, breaking the static-toolbox ceiling; RUBAS decomposes agent safety into four scoring dimensions and uses RL to balance helpfulness against safety. Together they address whether your agent system can be diagnosed, can self-extend, and can go to production safely — three checkpoints researchers tackled head-on today.
Three papers tackle 'how to make agentic AI work better' from three angles: the first (UIUC × Intel) profiles real agent workloads and finds the bottleneck is KV-cache management, not long prompts; the second (PwC) runs controlled experiments challenging the RAG-first default, showing grep often beats vector search in agent loops; the third (Microsoft Research) open-sources a complete agent training framework that lets the community train same-tier SOTA agents without relying on closed-source APIs.
Three papers, three angles on agent platforms: AgentFugue demonstrates that peer agents sharing a reasoning scratchpad can break through long-task collaboration bottlenecks; Can Agent Benchmarks Support Their Scores? reveals systematic flaws in current agent benchmark scoring mechanisms, urging us to re-examine leaderboard numbers; VibeServe lets agents auto-generate complete LLM serving stacks that outperform hand-tuned vLLM in niche deployment scenarios while matching it in standard ones. Together they answer: how can agents collaborate better, can we trust the evaluation numbers we rely on, and can agents build infrastructure for engineers?