CS224V splits task-agent evaluation into state updates and complete interaction: isolate the semantic parser, then test task completion, grounded queries, and valid actions with real users.
SkillEvo replaces single-turn QA evaluation with multi-turn interaction feedback so skill evolution doesn't stall after the first round, outperforming self-reflection by 23 points; SkillShapley brings Shapley values to skill step attribution — 99 evaluations approximate the exact ranking, revealing that 'decision-bridging steps' are the high-value ones; MindMemOS unifies memory management with an entity-property-time structure, hitting 94% on LOCOMO and lifting SpreadsheetBench success rate by 9.2 percentage points through skill evolution
Harness-IF reveals Coding Agent instruction following is overestimated by 3.6-7.4 pp because things the model would do anyway are counted as compliance; SHE decomposes the harness into four safety components and auto-evolves from trajectory failures, cutting ASR by 3.1x while improving correctness; SBCO uses a decomposed verifier bank with text gradients for harness self-improvement, matching Gödel Machine at 4-5.5x lower compute on planning tasks
Muscle Memory proposes 'compiled memory' over retrieval-based memory, winning 88.9% of personalization matchups across 90 scenarios; MoRSE uses role-subtask conditioned LoRA experts to significantly outperform prompt-only role differentiation in code generation; ASCon builds a unified failure attribution model, improving by 5.83%, 10.63%, and 14.73% across three attribution targets
OneDayAgent's decompose-remember-verify harness hits 0.821 new SOTA on AgentIF-OneDay and works unchanged across five backends; The Horizon Gap surveys 1,547 papers to find that six categories of long-horizon failure share a single structural pattern — outcome-only signals degrade as step count grows, driving the field toward denser process signals; Evo-Bench is the first benchmark for harness self-evolution — GPT-5.6 Sol peaks at +16.6 absolute gain, but Office tasks still need hand-crafted workflows
ToolLIFT lifts tool trajectories to function-level workflow graphs and consistently beats SOTA on three OOD benchmarks; SkillTV-Bench uses 681 cases to show skill-aware judge skills boost agent evaluation accuracy by 14.8pp; TRIO-20's prespecified equivalence study finds zero unauthorized calls from GPT-5.6 across 840 trajectories, but higher reasoning effort increases rule-probing rate by 14.3pp
Three papers examining AI Agent capabilities and limits from different angles: AutoMem shows memory management is a learnable skill — optimizing memory alone lifts a 32B open-source model to top commercial model levels; Shadow Evaluation tests whether frontier Agents can do open-ended AI research using real NeurIPS submissions — the answer is no, Agents can engineer but cannot research; Adaptive Adversaries reveals that existing safety benchmarks severely underestimate threats — adding adaptive multi-turn attackers jumps ASR from 0–1% to 14%. Together, these three papers deliver a sobering lesson: know where Agents can automatically improve, where they cannot, and that your security testing is probably insufficient.
Three papers tackle 'what goes wrong when agents hit production' from different angles: ProACT addresses when an agent should speak up in multi-user collaboration (an Agent UX design problem); the second uses real GitHub data to reveal that coding agents clash with their own PRs (a platform ops pain point); the third surveys five vulnerability classes of cyber-capable agents, using July 2026 HuggingFace/OpenAI incidents as case studies. Together, they form a crash course in post-deployment agent headaches.
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 today converge on one core question: **are AI Agents production-ready?** The answer is unanimously — far from it. HANDBOOK.md reveals that even the strongest frontier models achieve only **36.2%** SOP compliance when dropped into a simulated enterprise; a LangGraph paper delivers three actionable stateful workflow recipes plus a decision guide on when *not* to use LangGraph; and MM-ToolSandBox is the first benchmark to quantify how hard visually-grounded tool calling really is — the best of 12 models still falls below 50% success. Three dimensions — compliance evaluation, framework design, visual tool use — together map out exactly how far Agents are from real-world deployment.
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 today strike at the capability boundaries of AI coding agents from three angles: **ICAE-Bench** tackles interactive development under ambiguous requirements, exposing how current benchmarks lag behind the vibe-coding era; **EvoAgentBench** reveals the pitfalls of agent self-evolution ability transfer, where a mainstream method causes a −12.3 point negative transfer; **PERFOPT-Bench** opens the new track of performance optimization as an agentic task and finds that framework choice often matters more than model choice. The takeaway: production agent evaluation is far harder than existing tools suggest, and the field urgently needs benchmarks closer to real-world scenarios.
Three papers from ecosystem, failure, and memory angles: which open-source Agent frameworks are worth a long-term bet (beyond star counts), the six failure categories where Agents repeatedly stumble, and how to give Agents long-term memory that reasons across multiple entities. Together they form a 'framework selection guide + failure prevention checklist + memory system upgrade roadmap' for Agent platform developers.
Today's common theme: **the way we evaluate agents is itself broken**. The first paper audits major tool-calling benchmarks and finds nearly 20% of scores are wrong; the second uses replay analysis to show which benchmarks can be stopped early for reliable conclusions (SWE-bench is the exception); the third introduces the first multimodal web agent benchmark that jointly evaluates task completion and guide generation — screenshot input, dual-objective scoring, and even the strongest models complete less than 40%. Read all three for a complete picture of the crisis in agent evaluation and where to go from here.
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 examining real-world challenges for AI coding agents: the first systematically demonstrates how coding agents can be tricked into supply-chain attacks via manipulated READMEs, with defenses depending more on the harness than the model; the second introduces BPO, a reinforcement learning algorithm that branches only at high-entropy decision points for more efficient agent training; the third shows how MCP can serve as a standard protocol for connecting agents to domain-specific simulation tools in industrial settings like power grids, providing a replicable template for vertical-domain agent deployment.
Three papers tackling core agent platform pain points from different angles: the first proposes a framework for making e-commerce sites AI browser-agent friendly, boosting success rates from 49% to 89%; the second uses dynamic abstention-aware RL to teach search agents when to say 'I don't know'; the third introduces an agent OS for embodied robots whose multi-modal graph memory and context-isolated skill execution offer direct inspiration for general agent platforms. Together they cover the full chain from front-end UI design to inference reliability training to execution-layer memory architecture.
Three papers illuminate the AI agent landscape from very different angles: LHTB benchmarks 46 long-horizon terminal tasks and finds even the best model solves only ~28%; a second paper reveals a fragmentation effect in multi-agent systems that defeats per-agent monitoring; a third argues that in-process memory retrieval—1000× faster than cloud vector stores—fundamentally changes agent reasoning quality.
Three papers tackle AI Agent platforms from practical angles: the first exposes stealthy security threats in multi-agent systems and proposes activation-space detection of malicious agents (F1 +0.55 over graph methods in async settings); the second improves coding agent retrieval by introducing procedural similarity — finding code with similar solution steps rather than surface resemblance; the third is a wake-up call: the same LLM in different harnesses produces significantly divergent mid-task judgments, meaning harness design is never neutral.
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 converge on one question: how can Agent systems operate reliably? STRACE tackles noisy optimization inputs — precisely identifying root causes from massive noisy failure traces so automatic optimization stops getting derailed by redundant cases. The Blind Curator exposes an unsettling silent failure mode — the skill retirement mechanism in self-evolving Agents completely breaks down beyond a certain LLM judge bias threshold, and no amount of additional data can fix it. Severity Scale transforms 'how bad was this Agent attack' from binary success/failure into a seven-level action-harm score, finally giving security evaluation the granularity it needs. Read together: optimization quality, self-evolution soundness, security evaluation precision — three different layers, all pointing toward Agent trustworthiness.
Three papers today map the 'evolutionary frontier' of Agent platforms: EvoSOP lets agents extract reusable SOPs from past execution traces instead of replanning from scratch; AgenticSTS proposes a strict bounded-memory contract with five typed layers replacing endless context stacking; Spider 2.0-AIFunc reveals that AI functions are already embedded in cloud SQL syntax, yet the best model hits only ~67% accuracy — a new challenge every data agent must face. Together they outline three critical gaps agent platforms must close in 2026: tool efficiency, memory architecture, and data capabilities.
Three papers today converge on a single core issue: the massive gap between how AI Agent systems perform in idealized labs versus real-world deployments. AgentGym2 (ACL 2026) quantifies evaluation distortion with a new benchmark; an Agentic RL paper proposes engineering infrastructure for agents that self-evolve in production; and ComfyClaw demonstrates end-to-end skill self-evolution in image generation workflows. Read together, they form a complete map from evaluation → deployment → runtime evolution.
Three papers today attack the same core question from different angles: **how to make agent workflows truly reliable in production**. Mnemosyne brings the database Transaction concept into agent workflows, requiring every LLM output to pass admission control before taking effect. PaperPilot shows how to train a 9B model to plan multi-turn search workflows as DAGs and dynamically revise them based on user feedback. SEA lets agents self-improve on the fly while issuing auditable safety certificates. Together, the three papers nearly cover the full reliability stack for agent systems: execution-layer protection, training-layer workflow learning, and update-layer safe evolution.
Three papers tackling core agent platform pain points: ReContext offers a training-free inference-time fix so LLMs stop overlooking key evidence in 128K contexts; the second reveals systematic public-private divergence (3% → 40%) when agents debate across social hierarchies; the third raises alarms about three widely-cited coding agent benchmarks — only 8% of SWE-Perf tasks reproduce reliably.
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 three core Agent platform challenges: **upgrading memory from retrieval to reasoning state** (User as Code), **removing the central orchestrator while cutting costs** (DeLM), and **letting users quickly verify Web Agent results** (HANSEL). Together, they form a near-complete technical map for a high-trust Agent platform — memory layer, coordination layer, and explainability layer, each addressed by one paper.
Three papers spanning distinct dimensions of the AI Agent ecosystem: Qwen introduces the first Language World Model covering seven agent domains, enabling agents to train in simulated environments instead of relying on real APIs; Kuaishou's AgentX demonstrates industrial-scale multi-agent deployment, boosting recommendation algorithm iteration efficiency to 13.8x human output; OpenAI uses real Codex usage data to quantify how agentic AI is reshaping work across job functions, revealing that non-technical roles (legal, research) see even greater agentic dividends than engineers.
Three papers converge on one core question: **how do we actually evaluate whether an agent is good enough?** SWE-Explore isolates the most overlooked middle step of coding agents — understanding the codebase — and benchmarks it independently; Claw-SWE-Bench reveals that harness design (the adapter) is the real lever behind coding agent score jumps, with the same model leaping from 19% to 73% by swapping adapters; Red Queen Gödel Machine (Cambridge × NVIDIA) goes further by co-evolving the evaluator alongside the agent, breaking the ceiling of static benchmarks. Read together: **evaluation infrastructure is becoming the most critical competitive moat for agent platforms**.
Three papers tackling core Agent platform pain points: one decomposes Agent memory into four measurable system modules, revealing that current evaluations only checking 'did it get the answer right' are far from enough; one borrows the software engineering concept of 'design review' to enable automated verification of Agentic Workflows before deployment; and one uses 14 large-scale parallel experiments to prove that the benchmark leaderboard you trust reshuffles its rankings when the context changes — and proposes a more reliable alternative metric.
Three papers, three angles: **RigorBench** evaluates coding agents on process discipline rather than just pass rates, introducing five dimensions of engineering rigor; a production-focused paper shows how to customize and accelerate large multi-agent systems for enterprise use (4.48x throughput gain); and a governance paper proposes a formal protocol language for specifying human-agent boundaries in the SDLC — turning 'which decisions AI can make' from a line in a prompt into a machine-verifiable spec. Together they cover evaluation, deployment, and governance.
Three papers exploring the boundaries and breakthrough paths of agent capabilities. Sakana Fugu (Sakana AI) trained a 0.6B orchestrator model that learns to dynamically coordinate a pool of frontier LLMs, achieving public SOTA on SWE-Bench Pro and other benchmarks — the core thesis is that the orchestrator itself can be trained rather than hard-coded by engineers. NatureBench uses 90 real research tasks from Nature journals to ask: can coding agents actually make scientific discoveries? The best configuration only surpasses published SOTA by 17.8%, mainly by translating problems into familiar ML tasks rather than truly inventing new methods. Finally, Rising from the Ashes — six security researchers systematically map how agentic AI can take over five categories of labor-intensive tasks that have long plagued defenders, with 16 case studies as deployment references.
Three papers on agent platform infrastructure gaps: PlanBench-XL reveals top LLMs collapse under tool failure in large-scale ecosystems (GPT-5.4 drops from 52% to 11%); TU Munich provides the first technical taxonomy of 9 agent communication protocols (MCP/A2A/ACP/ANP) for principled selection; AMD's Arbor uses tree search as a shared cognition space for multi-agent collaboration, turning failures into useful exploration signals. Together, they outline three foundational infrastructure gaps in 2026 agent platforms.
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 paint a full picture of how agents land in the real world: Perplexity + Harvard Business School use production data to quantify the agent vs. chatbot gap for the first time — 87% faster task completion, and agents attract cognitively harder work; Self-Harness shows how agent scaffolding can automatically mine weaknesses and fix itself, yielding 33-60% relative gains across three models; The Consistency Illusion exposes a core trap in multi-agent debate — output-level consensus can mask fundamentally misaligned reasoning underneath. Read together, the signal is clear: an agent's real competitive edge isn't a stronger model — it's production-data-driven scaffolding self-improvement and rigorous validation of collective decision reliability.
Three papers challenging conventional wisdom in the agent space: ACCORD shows agents act on assumptions instead of observations and fixes it with active grounding (AppWorld 42% → 62.6%); 'The Illusion of Multi-Agent Advantage' proves auto-generated MAS underperforms single-agent CoT-SC at 10x the cost; 'Agentic Very Much' provides large-scale GitHub evidence that coding agent adoption in new projects has more than doubled year-over-year. Together they signal: agent tools are spreading fast, but the assumptions that 'multi-agent is always better' and 'agents understand your instructions' are being challenged by data.
Three papers targeting three critical infrastructure layers of Agent platforms: HarnessX introduces a 'harness as evolvable component' framework that turns static Agent scaffolding into a self-optimizing system (+14.5% average across 5 benchmarks); the second studies skill-conditional trust routing in multi-agent collaboration, revealing when fine-grained trust actually helps and how attackers can hijack it; OCELOT tackles security with a 'posterior leakage budget' mechanism to prevent Agents from gradually leaking user privacy to external services. Together they cover framework design, multi-agent governance, and privacy security — exactly the three pitfalls most commonly hit when shipping Agent platforms to production.
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 paint the 'agent reality of 2026': UC Berkeley's real-workplace benchmark shows top agents pass only 2.6% of the hardest tasks; Microsoft finds developers spontaneously develop 4 oversight behaviors that tools don't support; Reins AI argues task-level monitoring can't see the worst structural failures in early-stage agent systems.
Three papers tackle the same core question from different angles: **how to evaluate and operate AI Agents under real deployment conditions.** Emergence World builds a multi-agent sandbox that runs continuously for weeks, exposing behavioral drift and cross-model contamination invisible to short-term benchmarks; a survey paper establishes a complete taxonomy for agent environment design (8 attributes x 8 domains) and proposes symbolic vs. neural synthesis paradigms; Martin Monperrus's position paper declares outright that coding agents have crossed the threshold and human code review can retire.
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 converge on one theme — moving agents from experiments to reliable production: a multi-agent troubleshooting architecture deployed at hyperscale cloud with 90%+ autonomous resolution; a memory mechanism that lets agents learn from past tool-call successes and failures without retraining; and the first systematic comparison of six AI-assisted development process frameworks across six dimensions.
Today's three papers center on **security boundaries and capability optimization for coding agents**: SABER introduces the first executable-workspace benchmark and finds even the best models have 54%+ dangerous operation rates; the second paper has 100+ real developers collaborate with a secretly sabotaging AI agent for five hours — 94% never noticed; SePO shows that auto-optimizing system prompts alone (no model changes) yields an average 4.49-point gain across five benchmarks. Together they remind platform builders: agent safety is harder to measure and harder to catch than assumed, yet low-cost improvement paths exist.
Three papers mapping to three layers of the agent platform stack: AgentJet (training layer) introduces a distributed framework for simultaneous RL training of multiple heterogeneous LLMs, solving the fundamental limitation of single-model-only training tools; AdaPlanBench (evaluation layer) reveals with a 67.75% ceiling that LLM agents are far from ready for real-world scenarios where rules are disclosed progressively — it is the first benchmark to systematically quantify this adaptive planning capability; Beyond Tokens (communication layer) surveys multi-agent systems that replace text with embeddings for inter-agent communication, providing a taxonomy to evaluate the engineering trade-offs of this new communication path.
Three papers tackle agent infrastructure decisions: ADK Arena quantitatively compares LangGraph, AutoGen, CrewAI and other frameworks on real-task completion rates and costs; Agent Memory offers the first computer-systems taxonomy of 10 memory designs covering latency, bandwidth, and scalability trade-offs; Search-Time Contamination questions deep research agent benchmarks—agents can search for answers during evaluation, inflating scores by up to 4%. Together they provide new quantitative tools for three core platform decisions: framework selection, memory architecture, and evaluation trustworthiness.
Three papers on three deep agent-system questions: **memory architecture** (which design generalizes?), **self-evolution** (can AI build agents autonomously?), and **security blind spots** (how domain-dependent is CUA safety?). AutoMEM shows agents that actively manage their own memory generalize better than those relying on external pipelines; Meta-Agent Challenge reveals that frontier models still fall well short of autonomous agent development; Domain-Conditioned Safety finds Claude Sonnet 4.6 has 0% prompt-injection ASR on web tasks but 100% on code tasks — all three challenge core design assumptions in agent platforms.
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 tackling 'how to build more reliable, evolvable Agent systems' from different angles: the first reveals real LLM call costs in multi-model Agent systems through execution traces, giving platform engineers hard numbers; the second proposes treating the entire memory pipeline as self-evolving code to fix memory-architecture drift in long-running tasks; the third exposes evaluation blind spots in Agent continual learning benchmarks—current benchmarks can't tell whether agents actually learned anything—and introduces a more rigorous controlled stream framework.
Three papers at three different layers: BenchTrace ran 1,821 agent failure episodes and found GPT-4.1 and Qwen3-32B pass less than 30% on diagnosing their own failures — reflection is far weaker than assumed; Beyond Autonomy distills a three-tier governance architecture from enterprise SaaS production, filling the missing 'governance' piece in current agent frameworks; Insuring Every Action prices every agent action using actuarial concepts and introduces reserve capital budgets, creating an entirely new runtime risk vocabulary. The common thread: the core challenge of enterprise agent deployment has shifted from 'can it do the job' to 'what happens when it fails, who reviews it, and how do you quantify the damage.'
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?