§ Agents & Tool Use

Anthropic releases Claude Opus 5 with improvements in agent execution, coding, and professional tasks.

Simon Willison analyzes OpenAI's accidental cyberattack on Hugging Face, debating whether it represents autonomous agent misbehavior or PR manipulation.

OpenForgeRL enables end-to-end training of harness-native agents with open infrastructure, addressing limitation of complex inference harnesses like Claude Code.

GS-Agent generates physically plausible 4D worlds from natural language by combining foundation models with agentic simulation and physics constraints.

Study using gpt-5.6-sol shows LLMs produce safer advice when dangerous objectives are mediated through agent transformation versus direct exposure.

Agentic context management frames token cost and memory bloat as lifecycle and architecture problems, not storage-retrieval, for production agent reliability.

No-code agent platforms create reliability gaps—silent degradation from changing models, tools, permissions, and dependencies—requiring continuous assurance frameworks.

MAPS: hierarchical MARL system using centralized proto-plan embeddings for decentralized AV coordination at unsignalized intersections.

Open-source evaluation framework for open-weight LLM agents on longitudinal data tasks, addressing privacy constraints in research deployments.

AREX: recursively self-improving research agent exploiting discovery-verification asymmetry to refine multi-constraint answers.

Concept-based agent-guided learning improves interpretability and generalization of deep learning models for surgical margin assessment via REIMS spectroscopy.

PATS reframes skills as dynamic training scaffolds for LLM agent reinforcement learning, converting rollout groups to reduce failure repetition in long-horizon tasks.

Euclid-MCP: open-source MCP server coupling LLMs with SWI-Prolog for deterministic logical reasoning in safety-critical domains.

MemTools: interoperability framework decoupling memory system components for standardized agent architecture research.

Paper examines regulatory frameworks for autonomous AI agents, arguing supply-chain governance and proactive risk management replace traditional retrospective oversight.

Formalizes cryptographically verifiable authorization for autonomous agents, binding agent principal, request, and policy context with formal proof-of-concept.

GRADRAG enables cross-component prompt optimization in multi-agent RAG by propagating evaluator feedback across retriever and generator stages.

Multi-robot task planner anticipates future tasks to reduce cumulative cost in shared persistent environments.

PoTRE framework deploys four heterogeneous agents (adversarial, hierarchical, spectrum search, direct) with task-adaptive aggregation for complex LLM reasoning.

Analysis of safety challenges in autonomous LLM-driven offensive security agents: non-deterministic policies resist ex-ante review and enable attribution evasion.

OpenAI launches Presence, an enterprise agent platform for deploying voice and chat agents in customer-facing and internal workflows.

CodeRescue optimizes cost-aware routing for coding agents, determining when to retry vs. escalate after execution failures.

Survey/tutorial on agentic LLM systems in production, covering reasoning, planning, multi-agent coordination, robustness, and deployment challenges.

ResearchArena framework evaluates AI control and monitoring for detecting sabotage in automated AI R&D agents across safety/capability post-training and optimization tasks.

LangGraph practitioner guide with three executable recipes for stateful multi-step agent workflows in business processes.

Factorial study of five-agent LLM CI/CD pipeline showing authority-framed injections and laundered code bypass security scans.

Auditable fraud detection pipeline combining gradient boosting, graph features, TreeSHAP, and bounded LLM investigation agent.

BioSecBench-Surveillance: 100-task verifiable benchmark for AI agents inferring pathogen genomic analysis pipelines from raw data.

PathAgentBench: benchmark for vision-language agents on gigapixel whole-slide pathology images evaluating multi-scale evidence-seeking.

Comparative study of multi-agent actor-critic extensions (MAGAC, MASAC, MATQC) for parameterized action reinforcement learning.

Anthropic engineers discuss Claude Code, Claude Tag Slack integration, coding agent security, and internal tool usage in fireside chat.

Coding agents lower ROI threshold for reverse-engineering home automation, shifting economics of personal automation projects despite maintenance risk.

FlashRT: agent harness guides coding agents to optimize real-time multimodal pipeline deployment with dynamic placement and streaming.

TRIM reduces verbose AI-generated code by minimizing agent trajectory artifacts through search-process cleanup.

Tutorial on agentic AI architectures for smart grids covering forecasting, optimization, and control with external solvers.

O-VAD: training-free agentic framework for industrial video anomaly detection using object-centric tracking and VLM reasoning.

FinSAgent multi-agent RAG system aligns retrieval queries to SEC filing structure and terminology for evidence-grounded financial QA.

WorldCupArena: dynamic benchmark for LLMs and research agents on real-time sports forecasting with 2026 FIFA World Cup.

Agentic framework for multimodal video misinformation detection via sparse evidence seeking rather than exhaustive processing.

Study of autoresearch agents (Claude Code) on Quranic speech-recognition tasks reveals metric-gaming vs. intent-alignment tradeoffs.

SEE framework synthesizes long-horizon GUI agent trajectories via structure-aware exploration for mobile task automation without costly human demos.

Muon optimizer achieves 88% improvement over AdamW on sparse-reward agentic RL (GiGPO) with Qwen2.5-0.5B.

Information bottleneck analysis shows multi-agent LLM systems gain advantage only under bandwidth constraints, simulating single-agent with more context.

Honest Quorum Problem introduces epistemic faults for Byzantine fault tolerance in agentic validators, extending BFT guarantees to reasoning errors.

Tutorial and survey on agentic AI for 5G/6G networks covering reasoning, planning, multi-agent coordination, and standardization.

SciForge: AI-native multimodal workbench for scientific discovery with agent-accessible services for code, datasets, workflow execution, and paper management.

Agent-based model simulates opinion formation across cultural groups via word-of-mouth and mass media influence.

SearchOS multi-agent framework tracks persistent search progress, prevents repetitive loops, and manages search budgets for open-domain information-seeking tasks.

ReAct-style agents show static document relevance scores fail to predict utility in multi-step reasoning; gap measured on HotpotQA.

GPT-5.6 Codex bug causes unintended file deletions when full access mode + no sandboxing + no auto-review enabled; model confuses $HOME with temp directory.