AI Agents

PANINI-DR decomposes complex investigations into connected subquestions and stores findings, dependencies, and evidence in a shared semantic workspace. Its evaluation suggests that planning, persistent state, and claim-level verification can substantially improve deep-research reliability.

A new study examines how desktop software evolution affects the performance of language-model agents and introduces ANIS, a framework for synthesizing agent-native interfaces. The results suggest that better tool design can improve task success while sharply reducing online token usage.

ClawLoop connects task synthesis, model training, and capability diagnosis into a recurring pipeline for agents operating in sandbox workspaces. The approach aims to reduce reliance on external teacher models and one-shot human-designed data.

Alibaba researchers introduce TerminalHorizon, an automated pipeline for synthesizing long-horizon terminal tasks that preserve exploration, implementation, verification, and iterative improvement. The resulting data is used to fine-tune Qwen3.5-27B into TerminalHorizon-27B.

Work2Skill treats human desktop demonstrations as more than action traces, extracting workflow rules, step dependencies, and completion checks for computer-use agents.

SoftwareGym is an automated pipeline for creating runnable GUI applications that can be used to train and evaluate computer-use agents. Its synthesized environments improved a model’s out-of-domain performance on OSWorld.

CASE moves medical AI beyond answering questions over preselected records by letting agents decide what clinical and MRI evidence to inspect. The work combines an evidence-seeking environment, a longitudinal benchmark, and agentic post-training for compact vision-language models.

Stanford researchers propose Stateless Language Agents, moving durable research state out of agent conversations and rebuilding focused contexts for every call.

Meta’s Muse and OpenAI’s Dots are competing to make privacy a defining feature of AI agents. Yet isolated environments, user controls, and real-world behavior reveal a widening gap between promises and proof.

Anthropic has moved Claude Managed Agents dynamic workflows into public testing. The system uses a lead agent to create a plan, delegates stages to multiple agents, and combines their outputs for large workloads.

SAP researchers introduce Synthesis Through Simulation, or STS, a schema-free approach that lets an LLM agent generate enterprise data by operating through policy-enforcing APIs in simulated business environments.

AgentGarten combines simulators, game engines, and a shared neural renderer to give agents interactive worlds with persistent rules and richer visual observations. Its playbook mechanism lets later agents inherit and refine skills learned in earlier rounds.

Long-running agents generate too much fragmented activity for users to inspect step by step. A new study introduces EBG, a training-free behavior graph that connects consequential decisions with evidence from the agent’s original execution context.

WebFovea finished second in the 2026 WebRetriever Challenge and argues that reliable web agents need more than capable multimodal reasoning. The system improved the same model’s hidden-set score from 31.0 to 57.0 by hardening the interaction harness.

GUI-HARVEST improves the runtime harness around a GUI model while keeping the model weights frozen. It links actions to visual state changes, compares repeated runs, and turns verified failure patterns into reusable code updates.

A new study formalizes personal-agent mediated recommendation, where an agent uses authorized cross-platform history to revise a platform’s initial ranking. It also introduces MediateRec and PAMO to study when intervention helps and when it causes harmful overrides.

MiniCorp is an office simulator that links internal agent collaboration with an evolving market. It aims to produce longitudinal and counterfactual enterprise data that real organizations rarely make available.

Web agents need to read third-party pages, but those pages can also carry instructions that hijack the user’s task. AdvSim2Real co-evolves tasks, attacks, and the agent inside a frozen web world model to improve both capability and robustness.

An agent can reach the correct end state without having established the evidence that should justify its actions. SafeActBench shows that the weakest links often appear before execution and inside dependency-heavy workflows.

ASCENT introduces an online test-time training method that distills verified execution experience instead of blindly imitating a single rollout. The approach improves success rates on ALFWorld and WebShop while reducing the number of turns required per episode.

Proactive LLM agents aim to turn idle computation into useful assistance before a user asks. A new paper frames the problem around task capability, temporal allocation, and trust, and introduces Proactivity-Gym for studying their long-term effects.