Agents

a quiet day lets us find some connections among recent themes

Moonlight & Mayhem (Raccoon Heist by Codex + GPT-5.6 Sol Ultra) On Wednesday I wrote about One-shotting a Raccoon Heist game using Claude Fable 5 , where I had Claude Fable 5 build a full working game from a premise I generated with GPT-3 and DALL-E four years ago . I decided to pose the exact same prompt to Codex Desktop running GPT-5.6 Sol Ultra - the mode where Sol makes aggressive use of sub-agents - to see how it would do. It produced a much better game! Here's Moonlight & Mayhem - GitHub repository here , including the textures and prompts it generated using gpt-image-2 . Your browser d...

Cloudflare has introduced Kitesurf, a cloud-hosted browser designed for AI agents instead of people. The company says the browser uses less computing power than Chromium for common automation tasks, helping developers build browser-based AI agents more efficiently.

Tool use transforms LLMs into agents that act beyond their training data, and for code-capable models, programmatic tool calling extends this further by replacing rigid JSON calls with scripts that chain and parallelize naturally. However, a systematic evaluation of tools as code on an established benchmark across current and prior model generations under real-world task conditions has not been conducted. In this work, we empirically compare programmatic tool calling (PTC) to native JSON tool calling across 14 language models on BFCL v4. In the programmatic tool calling paradigm, tools are ex...

Deciding which of two agents is stronger means playing games until skill outweighs luck, and every game costs money, model inference, or expert time. Since the number of games needed is unknown, fixed-budget evaluations either keep paying after the result is settled or stop before the agents can be told apart, while naive optional stopping with an ordinary confidence interval invalidates the stated level. We make such an evaluation stop as soon as its evidence suffices, with the guarantee intact. The Action-Informed Value Assessment Tool (AIVAT) reduces variance in imperfect-information games...

We give a formal mechanism design model for the continuous participatory governance of a deployed AI agent. The mechanism is built on the principle that governance should control an AI agent through resource allocation so as to make authorization self enforcing via compute budgets. The mechanism seeks to establish the Safe AI paradigm that compute is an effective governance lever. We situate our work as a compliance or commons overlay on a deployer. One governance period is an extensive form game in which verified human stakeholders arrive sequentially and contribute, on a provision or a reje...

Training terminal agents requires executable and verifiable tasks that are not merely solvable, but appropriately challenging for learning. Executable validation establishes feasibility, yet does not reveal how a task behaves relative to a given solver setting. In this paper, we present CalibForge, an autonomous terminal-task synthesis system that uses verified solver behavior to revise candidate tasks through adversarial solver calibration. Multi-solver calibration targets disagreement within a heterogeneous solver pool, whereas contrastive solver calibration targets a designated strong-pass...

Task-oriented conversational agents are evaluated using curated or automatically generated benchmarks, yet benchmark quality is rarely assessed. Poor benchmarks may contain inconsistent tasks, simplistic scenarios, or limited policy coverage, leading to unreliable evaluations. We introduce a reference-free framework that uses LLM judges to assess benchmark consistency, complexity, and policy coverage, while providing actionable diagnostics of weaknesses. We validate the framework by demonstrating agreement with independent human annotations and by evaluating benchmarks generated by LLMs of va...

Synthetic clinical benchmarks for enterprise AI agents can pass existing utility checks and still remain structurally unrealistic, especially in privacy-sensitive healthcare settings where operational data are hard to access. We study how to improve such benchmarks without breaking the downstream utility checks already used in practice. We formulate benchmark revision as utility-constrained realism improvement: dataset changes should increase realism while staying above an operational utility floor. We instantiate this idea on a care-gap benchmark derived from Synthea-generated patients exerc...

Despite advances in artificial intelligence (AI) across multiple sectors, today's AI tools, including deep learning and generative AI, still fail when embedded into physical systems, such as robots and vehicles operating under real-world physical laws. This stems from their inability to maintain reliable world models for long-horizon planning under uncertainty and generalize to unseen scenarios. In this context, wireless networks, through pervasive sensing and communication, can orchestrate physical intelligence. However, current architectures optimize throughput, latency, and reliability and...

Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground. We introduce EnvACE, an agentic reinforcement learning method that replaces external environment interaction during training with world rehearsal. The policy alternates between acting and rehearsal: it first generates a tool call, then plays the role of the environment to produce the response induced by that action, and conditions subsequent decis...

Agents backed by large skill libraries must decide which skills to load and in what order. Loading the entire library into context is expensive and provides no structure for autonomous sequencing. We study two systems for this problem over a corpus of 690 skills: a hybrid ranker combining lexical and dense-embedding retrieval for sparse, on-demand loading, and a typed knowledge graph encoding workflow relations such as prerequisites, data flow, and ordering. On a set of 117 realistic, non-echoing queries, the hybrid ranker retrieves the correct skill within the top five in 73.5% +/- 8.0 of ca...

Web agents observe a browser through text, pixels, or both, and the choice is usually fixed once for all tasks. We measure six observation modes across eight site-model combinations (cells) on VisualWebArena and WebArena and ask what choosing per task would buy. The modes are complementary: each solves tasks the others miss, they fail in structurally different ways, and the best choice reverses between task sets. The obvious prize, an oracle that picks a winning mode for every task, looks large but is inflated by run-to-run noise: rerunning the same mode on the same tasks changes 12-14% of ou...

Incident Report: unsanctioned agent behaviour during cyber testing It happened again . This time it was the UK government's AI Security Institute who accidentally attacked other companies while running an evaluation with models with the safety filters turned off. From their technical paper (PDF): During a cyber evaluation, from 25 to 28 July 2026, AI agents engaged in sustained, unsanctioned activity directed at what were, in practice, real people and organisations. These attempts were unsuccessful and, to the best of our knowledge, no real-world harm resulted. [...] Across 122 evaluation att...

The serial entrepreneur joins the e-commerce company as CPO to lead its AI agents.

In partially observable reinforcement learning, agents face a dual bottleneck: they must explore to encounter rewarding states and retain that experience in memory to optimize their policies. Exploration bonuses and memory architectures are traditionally evaluated in isolation, leaving their interaction unmeasured, and standard notions of sparse reward conflate temporal signal density with what the reward actually supervises. We present a controlled study crossing episodic exploration bonuses with diverse neural memory architectures across three environments that vary how the content of memor...

AI-supported care planning can help clinicians, patients, caregivers, and care teams coordinate complex decisions across clinical, functional, psychosocial, and environmental needs. However, many AI systems present recommendations as fixed outputs, limiting stakeholders' ability to inspect, challenge, and revise plans when they conflict with clinical judgment, patient values, or real-world feasibility. We present CoPlan - a Co-Intelligent and Contestable Interface for Human-AI Care Planning. CoPlan uses a multi-agent workflow in which specialized AI agents generate candidate interventions and...

Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agents typically treat all steps within a trajectory uniformly during both supervised fine-tuning (SFT) and reinforcement learning (RL), failing to distinguish useful actions from erroneous or redundant ones. In this paper, we propose Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outc...

Agents for long term reasoning require a memory that can be efficiently and effectively updated over time, as new facts and external feedback continue to arrive. Recently, graph memory has been adopted to offer structural organization for multi-hop retrieval and reasoning. However, existing methods store all memories in a flat graph, and accumulated historical memories can introduce irrelevant contexts and increase the cost of evidence selection during retrieval. Moreover, they typically update memory units independently, requiring repeated unit-wise rewrite to cover related changes. To addre...

Yet more rogue AI agents from OpenAI and Anthropic have been caught attempting to hack real targets online without permission. The discoveries add to a growing list of previously unknown incidents that have alarmed AI safety experts and intensified pressure for greater oversight of frontier systems. According to a report from the UK's AI Security Institute, which evaluates frontier models from top AI labs before they are released, agents powered by OpenAI's GPT-5.6-Sol and Anthropic's Mythos 5 went "engaged in sustained, potentially harmful activity directed at real people and organisations."...

Training LLM agents commonly relies on supervised fine-tuning from expert trajectories or online reinforcement learning over human-specified tasks with handcrafted verifiers. Though effective, both remain bottlenecked by externally specified tasks and supervision signals, limiting the scalability and diversity of agent training. We study an environment learning paradigm in which agents acquire interaction and manipulation capabilities solely through environment interaction, without externally specified tasks. We propose State2State, an environment-derived mid-training method that converts exp...

The week-old Open Secure AI Alliance, spearheaded by Nvidia and grown to over 120 companies, already has proposals out for defending against AI agents.

SocietyBench evaluates LLM agents on forecasting real social-world events by ingesting news/social media timelines and measuring counterfactual event prediction.

PAST-Bench benchmarks recursive self-improvement in personal AI agents by measuring whether retained preferences, task histories, and skills improve performance over sessions.

Video-DeepResearch extends multimodal agents to video streams; identifies modality bias and knowledge leakage bottlenecks in current models.

HIVE benchmark reveals voice transcription and keyboard perturbations reduce LLM accuracy across instruction tasks.

Game-theoretic analysis shows foundation model agents achieve rational cooperation through similarity inference, challenging decoupled-agency assumptions.

ContinualSkillBench evaluates whether LLM agents can continually acquire and reuse skills across 500 interconnected task domains.

MAFIA: query-only memory poisoning attacks on audited LLM agents via probing and factual injection.

Multi-agent clinical committees using Gemini show vulnerability to social shortcut cascades where peer consensus propagates errors across agents.

SAT-Edge-Agent deploys edge-based LLM agents on satellites for onboard intelligence under communication/power constraints.

AntiSkillBench: end-to-end benchmark evaluating privacy leakage, impersonation risk, and defenses in persona-skill pipelines for agents.

TARL: memory state framework for long-term agents mapping statements to five executable actions (add, ignore, revise, reject, defer) instead of binary write/hold.

Offline reinforcement learning framework trained on 31.7k records to predict oncology clinical trial portfolios under uncertainty.

DiagChain diagnostic benchmark evaluates LLM agents on staged attack chain reconstruction from security telemetry across 69 scenarios.

Apple’s long-awaited AI overhaul finally makes Siri the assistant it was always supposed to be. But after years of delays, the launch lands in an AI landscape where chatbots have evolved into agents that can code, reason, create media, and complete complex tasks. Siri AI is genuinely useful, yet it arrives at a moment when simply being a capable AI assistant no longer feels revolutionary.

LiveMem maintains persistent memory state in long-running LLM agents by decoupling working context from intrinsic memory lifecycle.

RoMeRL addresses memory-reward trap in self-evolving agents using reduced-order utility states to balance feedback coverage and irrelevant experience pollution.

SWE-Touch benchmark tests coding agents in shared workspaces with user code edits, exposing weaknesses in handling conflicting modifications during task execution.

Microsecond-cost anomaly detection for LLM agent failures using one-class echo-state networks trained only on healthy runs; tested across Qwen, Llama, and Gemini agents.

ParEvalLayer detects biased partial evaluations of LLM agents, preventing premature benchmark conclusions from incomplete task runs.

Digital Twin-Enhanced Multiscale Planning automates incident response via decision-theoretic agents, bridging abstract models to operational systems.

Storage is an active part of every agentic AI workflow. As agents retrieve enterprise knowledge, access persistent memory, reuse key-value (KV) cache data,... Storage is an active part of every agentic AI workflow. As agents retrieve enterprise knowledge, access persistent memory, reuse key-value (KV) cache data, execute tools, and generate new results, storage systems must continuously supply and preserve the data that moves the agent reasoning loop. Each agent step can trigger multiple storage operations, and those operations can repeat across… Source

Google and Kaggle launched a free 353,000-person course on AI agents using Gemma, focused on building and deploying agent systems.

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. When two OpenAI models hacked into the website Hugging Face in July, they weren’t trying to make money or commit sabotage—they were just looking for answers…

CompressAgent benchmark evaluates reliability of compressed agent control contexts across Qwen models and task families.

Wix Helpmate deploys deterministic executability gating to filter skill selection in LLM agents by account state feasibility.

Temporal replay framework evaluates enterprise agents against dynamic data across multiple moments within an episode, not just final state.

Search-GRT: RL method to train LLM search agents for multi-hop QA with guided retrieval to reduce sparse-reward training issues.

PROGRESS trains search-augmented LLM agents using coverage-guided RL rewards to improve query decomposition over outcome-only supervision.

TrajWiki proposes trajectory-based external memory for long-horizon dialogue agents with traceable, updatable, diagnostically transparent storage.

PMMC compiles multimodal memory at consolidation time for LVLM agents to preserve image-text binding and temporal updates without query-time overhead.

Neuro-symbolic governance framework for verifiable AI agents in decentralized digital twin ecosystems with semantic profile layers.

LAND model simulates 314K heterogeneous agents and human actors over 30 days to study emergent social dynamics via LLM-enabled ABM.

OpenAI has reportedly found evidence of additional agent misbehavior as it looks into the incident that occurred with Hugging Face.

AgentHPOBench evaluates LLM agents on sequential hyperparameter optimization across 30 ML tasks, assessing experimental interpretation and adaptive decisions.

SESA framework combines self-play curriculum learning with evolving procedural memory to distill failures into reusable skills for search agents.

Analytic memory abstraction for multimodal agents enabling filtering, aggregation, and temporal reasoning over accumulated observations.

Zero-Mem: Zero-token memory operations for LLM agents using encoder computation instead of LLM calls to reduce latency and token costs.

Knowledge workers are increasingly integrating AI agents into their workflows. Agents that function as "digital coworkers" offer clear benefits. For example,... Knowledge workers are increasingly integrating AI agents into their workflows. Agents that function as “digital coworkers” offer clear benefits. For example, they can review a bug report, implement and test a fix, push a patch, and ping a human for review. By handling routine tasks, agents have the potential to deliver large productivity gains. On the other hand, connecting a large language model… Source

Empirical study of inference-time scaling strategies for local computer-use agents under hardware constraints across multiple dimensions.

ORCA-bench evaluates LLM agents on production oncall root-cause analysis using real telemetry (Prometheus, Jaeger, OpenSearch) and code in realistic incident scenarios.

CS-RNR method enables agents in imperfect-information games to safely exploit flawed opponents with provable certificates on deployed strategy.

The deal gives Okta identity threat detection capabilities as enterprises seek to secure AI agents and other non-human identities across cloud environments.

CARP reputation-penalty mechanism prevents LLM agents from fabricating product listings using complaint signals without access to ground truth.

Budget-constrained human audit allocation for N LLM agents identifies miscalibration threshold where confidence-ranking underperforms random selection.

MemHarness reconstructs retrieved memories contextually rather than replaying them verbatim, reducing negative transfer in LLM agents.

EMBL AI Librarian provides structured knowledge layer for AI agents querying 40M+ life-sciences records from Europe PMC.

Qwen-UI-Agent technical report describes cross-platform GUI agent for mobile, web, CLI, and long-horizon task automation.

ParliamentBench evaluates deceptive reasoning in 16 LLMs via Secret Hitler game framework with 1,600 adversarial matches.

AI engineers adopt ontologies to constrain probabilistic agents within deterministic logical boundaries, reviving semantic web techniques.

As Meta pours billions into AI infrastructure and agents, Zuckerberg is working to convince investors that the payoff will be worth the price.

On the company’s second-quarter earnings call Wednesday, CEO Mark Zuckerberg said Meta sees a “large enterprise opportunity” spanning AI agents, APIs, compute, and internal software.

Meta is all-in on AI, and sometime soon, the company is going to make a big push into personal AI agents that can do things on your behalf. On Wednesday's Q2 2026 earnings call, CEO Mark Zuckerberg previewed a high-level vision of how the company is thinking about personal agents and what it will do to make them viable for users - and how it will convince less technical people to give them a shot: Soon we will have agents that can work 24/7 on your behalf to help you achieve your goals and improve your life, your health, your relationships, your finances, whatever you want. The first domain t...

Study evaluates AI agents on open-ended research tasks graded by original paper authors, providing evidence for AI R&D automation feasibility.

OmegaUse-OfficeVal benchmark evaluates LLM agents on 100 long-horizon office-suite tasks with cost grounding and human labor baselines.

Cost-aware stopping mechanism (CAM-DF) for LLM agent tool acquisition balancing task coverage against cost, context load, and privacy.

Setoka benchmark evaluates hierarchical user understanding in memory-augmented personalized agents beyond explicit fact retrieval.

AgentSnare uses adaptive deceptive observations to mislead LLM-based penetration testing agents, defeating static artifact recognition.

The startup analyzes calls, messages and CRM data to identify effective sales techniques and turn them into playbooks for AI agents.

TREK benchmark evaluates LLM agents on complex, executable travel itinerary planning with verifiable constraints.

Three-class detection framework distinguishes humans, bots, and AI agents in browser automation traffic; binary classifiers misclassify 39.1% of agents as human.

Two-call self-refinement outperforms five-agent pipeline on Qwen2.5-7B; multi-agent systems suffer error accumulation, dropping GSM8K accuracy to 45% with JSON format.

TSDS framework deploys ReAct agents at edge via convergence probe for reasoning budget and perplexity-based deferral to cloud model.

The deal is Cyera's third acquisition this year.

Desktop-Delta Bench isolates whether computer-use agents understand GUI state transitions caused by actions, beyond end-task success metrics.

UniMem hybrid episodic-parametric memory for LLM agents resolves stability-plasticity dilemma across boundary-agnostic evolving task streams.

MemLens provides value-aware memory management for LLM agents with interactive analytics, treating memory records as first-class objects to reduce redundancy.

OpenAI field report documents how AI coding agents accelerate scientific computing workflows in genomics and adjacent domains.

Google ships Gemini API Managed Agents with Gemini 3.6 Flash and tool-use hooks for production agent deployment.

Messier consolidates 957K evaluation records across 30 benchmarks and 714 agents to enable comparable cross-benchmark agent assessment.

Custom harness engineering distributes security controls for AI coding agents without vendor lock-in, enabling organizational scaling.

RSIBench-Data isolates LLM agents' data-centric research capability in post-training loops, decoupling research from systems optimization.

OpenAI's Akshay Nathan details ChatGPT Work product strategy: Sites, memory, subagents, finance, no-code tools scaling from 0 to 10M users.

HiSkill hierarchical skill graph framework organizes LLM agent trajectories into structured graphs linking high-level skills to atomic operations for long-horizon tasks.

Joint agent-speculator RL aligns next tool-call prediction with deployed agent behavior by unifying speculator and agent in single model, reducing latency.

WorkSurface-Bench evaluates enterprise agents on multi-source knowledge routing across documents, tables, and graphs with 1,151 tasks.

HYSET evaluates tool sets jointly for LLM agents via hyperedge prediction, improving on sequential tool retrieval.

Perplexity has expanded its agentic Personal Computer tool to Windows, allowing computers running the world's most popular OS to be used as a locally run AI system. Like the Mac version that Perplexity launched in April, Personal Computer for Windows operates like a "general-purpose digital worker" that can access local files and apps to perform actions on your behalf, such as creating documents and updating spreadsheets. This launch builds on Personal Computer integrations that Perplexity launched for Microsoft's 365 workspace apps and Teams virtual meeting software in May. Personal Computer...