The Archive

Recent advances in large language models (LLMs) have facilitated the widespread deployment of LLMs as interactive agents capable of reasoning, planning, and tool use. Despite strong performance on existing benchmarks, such agents often exhibit notable degradation when deployed in real-world settings, where environments are inherently stochastic and imperfect. We argue that this discrepancy arises from a fundamental mismatch between idealized training settings and real-world interaction dynamics, where current paradigms rely on carefully curated task instructions and stable, well-controlled en...

Microsoft Copilot Cowork Exfiltrates Files The biggest challenge in designing agentic systems continues to be preventing them from enabling attackers to exfiltrate data. In this case Microsoft Copilot Cowork (yes, that's a real product name ) was allowing agents to send emails to the user's own inbox without approval... but those messages were then displayed in a way that could leak data to an attacker via rendered images: Because these messages can contain external images that trigger network requests to external websites, data can be exfiltrated when a user opens a compromised message sent ...

Amid rapidly growing adoption of enterprise-level AI agents, there’s a disconnect emerging between ambition and execution. Although 85% of organizations say they want to be agentic within the next three years, 76% say their current operations and infrastructure can’t support that change. They cite a lack of readiness across people, processes, and workflows. The sticky…

Today, I’m talking with Google and Alphabet CEO Sundar Pichai, in a conversation we recorded just after the Google I/O developer conference. This is the fifth year Sundar and I have sat down after I/O, and it’s become one of my favorite Decoder traditions. There’s always a lot of news at I/O, and this year was no exception — Google has powerful new Gemini models, it’s putting AI agents in everything, and it’s making huge changes to Search on both the web and YouTube that will once again reshape the information ecosystem. That’s a lot to talk about, and Sundar and I got into all of it. But I a...

MobileGym enables parallel RL training for mobile GUI agents via deterministic JSON-based state representation, supporting hundreds of concurrent instances.

DiscoverPhysics benchmark tests LLM reasoning by having agents discover physics laws in simulated worlds with non-standard dynamics.

Claw-Anything benchmark evaluates LLM agents as always-on assistants with access to long-horizon histories and interdependent backend services.

VeriTrace proposes explicit feedback mechanisms for research agents to evolve mental models and prevent error propagation in uncertain information.

Auto Benchmark Audit uses agentic framework to systematically audit AI benchmarks for hidden dependencies, specification gaps, and grading flaws.

CausaLab environment benchmarks LLM agents on interactive causal discovery with validation of both solutions and underlying causal mechanisms.

The nine-year-old startup is replacing hundreds of employees with thousands of AI agents.

68-cell empirical study: LLM agents show +19.69pp higher sensitivity to semantic noise vs. surface noise across reasoning tasks.

ProAct: proactive agent architecture using idle-time compute to predict and prepare for future user requests via dialogue history analysis.

Practitioner observes that phrasing dependency relationships improves agent instruction compliance from 75% to 90%+ across email/Slack/Sheets automation.

Hugging Face revives PapersWithCode.co with new SOTA tracking features for agents, vision, and time-series; week 1 update.

- NEW: Tool Description: Workflow — Describes the Workflow tool for opt-in deterministic multi-subagent orchestration, including script metadata, agent hooks with plain-text or structured returns, pipeline vs. parallel control flow, token budgeting, quality patterns, concurrency limits, and resume behavior. - NEW: Agent Prompt: Workflow subagent plain text output — Instructs workflow-spawned subagents to return raw final text as the calling script's parsed value, avoiding human-facing confirmations, markdown wrappers, or SendUserMessage delivery. - NEW: Agent Prompt: Workflow subagent structu...

Multi-agent LLM framework with Creator and validation agents generates physics-constrained constitutive models ensuring compliance with continuum mechanics laws.

OpenSkillEval automated framework audits open-source skill-augmented LLM agent systems for quality, compatibility, and cost-performance trade-offs.

PCSP shared RL policy for persona-consistent NPC control in life sims achieves 22x faster inference than LLM baselines on 300-persona benchmark.

Co-ReAct integrates step-level rubrics to guide ReAct agent reasoning in multi-step search tasks.

PushBench evaluates quantitative goal persistence in long-horizon LLM agents via work-unit completion.

Goal-conditioned RL agent leverages parallel all-goals learning by jointly predicting values and actions for every goal simultaneously.

OpenAI named Leader in Gartner's 2026 Magic Quadrant for Enterprise AI Coding Agents; Codex cited for innovation and scale.

Daytona CEO discusses agent infrastructure platform achieving 74% MoM growth, 850K daily runs, and bare metal sandboxes for RL evaluation.

Anthropic's June 15 Agent SDK pricing changes signal shift toward managed agents, pressuring third-party integrations and local deployment strategies.

Exploration is a prerequisite for learning useful behaviors in sparse-reward, long-horizon tasks, particularly within 3D environments. Curiosity-driven reinforcement learning addresses this via intrinsic rewards derived from the mismatch between the agent's predictive model of the world and reality. However, translating this intrinsic motivation to complex, photorealistic environments remains difficult, as agents can become trapped in local loops and receive fresh rewards for revisiting forgotten states. In this work, we demonstrate that this failure stems from a lack of spatial persistence a...

Autonomous agentic systems are largely static after deployment: they do not learn from user interactions, and recurring failures persist until the next human-driven update ships a fix. Self-evolving agents have emerged in response, but all confine evolution to text-mutable artifacts -- skill files, prompt configurations, memory schemas, workflow graphs -- and leave the agent harness untouched. Since routing, hook ordering, state invariants, and dispatch live in code rather than in any text artifact, an entire class of structural failure is physically unreachable from the text layer. We argue ...

Large language model (LLM)-based multi-agent systems increasingly rely on intermediate communication to coordinate complex tasks. While most existing systems communicate through natural language, recent work shows that latent communication, particularly through transformer key-value (KV) caches, can improve efficiency and preserve richer task-relevant information. However, KV caches also encode contextual inputs, intermediate reasoning states, and agent-specific information, creating an opaque channel through which sensitive content may propagate across agents without explicit textual disclos...

LLM-powered AI agents require high-frequency state exploration (e.g., test-time tree search and reinforcement learning), relying on rapid checkpoint and rollback (C/R) of the complete sandbox state, including files and process state (e.g., memory, contexts, etc.). Existing mechanisms duplicate the entire state, causing hundreds of milliseconds to seconds of latency per C/R, which severely bottlenecks deep search and large-scale fan-outs. This paper observes that subsequent checkpoints in AI agents are highly similar. Therefore, instead of full duplication, a sandbox should only duplicate the ...

Autonomous systems have achieved superhuman performance in isolation or simulation, yet they remain brittle in shared, dynamic real-world spaces. This failure stems from the dominant single-agent paradigm for physical applications, where other actors are ignored or treated as environmental noise, preventing effective coordination. Here we show that multi-agent reinforcement learning provides the essential safety scaffolding required for real-world interaction. Using high-speed quadrotor racing as a high-stakes testbed, we train agents to navigate complex aerodynamic interactions and strategic...