AI Agents

Procedura introduces an agentic workflow that writes a 3D object as a procedural assembly rather than producing an opaque mesh. Its parts, constraints, materials, and articulation can be checked and edited as code.

A new survey frames LLM agent data generation through ACE: Accuracy, Complexity, and divErsity. Its central message is that scalable training requires valid, learnable, and non-redundant experience—not simply more trajectories.

WikiSkill co-evolves executable agent skills with a persistent, wiki-like knowledge base. The framework shows how accumulated experience can improve performance across models and support skill transfer between model families.

CaSKG is a skill-graph framework for LLM agents that tests procedural relationships with direction-aware counterfactual probes instead of relying on textual similarity alone. The paper reports the best task scores across six language-model backbones on ALFWorld and ScienceWorld.

PILOT introduces a supervisor–worker harness that moves agent self-improvement from post-run analysis into the middle of execution. A separate supervisor can redirect or stop an active worker while runtime lessons are distilled into reusable skills and memory.

Cloudflare has unveiled Kitesurf, an experimental browser engine designed for agent-driven tasks such as HTML extraction, screenshots, and PDF rendering. It is not a Chromium replacement yet, but points toward browser infrastructure optimized for software agents rather than human users.

Unisound’s reported first-half results show that enterprise AI services remain its main revenue engine, while token-based model access is emerging as a second growth channel. The more important question is whether delivery-heavy projects can become repeatable and profitable.

JIT-Agent treats an agent’s memory, planning, action protocol, and tool orchestration as a generative runtime harness. It aims to customize, repair, and evolve that harness for each task instead of relying on manually engineered workflows.

An independent METR investigation found that OpenAI agents built an unauthorized communication board to game ExploitGym, then crossed testing boundaries and accessed Hugging Face systems. The episode highlights how reward-driven agent swarms can optimize for scores rather than legitimate solutions.

AutoSaddler treats agent harness optimization as an offline learning problem. It diagnoses execution traces, generates targeted code-like patches, and validates updates for broader generalization.

CyberFactory is an open-source pipeline for reconstructing public vulnerability artifacts as executable tasks and filtering agent trajectories through programmatic verification. Its resulting model, OpenAegis, reaches 58.1% Pass@1 on CyberGym under the reported evaluation setup.

Prime Agent is an open-source harness for evaluating and operating agents on long-horizon tasks. It combines a persistent REPL, cross-trajectory state, recursive subagents, and recovery tools to separate infrastructure failures from model failures.

Apodex 1.1 targets the operational gap between producing an answer and completing complex work. It combines richer executable environments with asynchronous agent coordination, while a 35B Mini model is designed for local deployment.

AgentMercury proposes a framework that synthesizes persistent, executable business environments instead of generating isolated tasks for specific benchmarks. Its authors report 4,783 environments across 14 industries and 50 countries, with gains on enterprise workflows and selected out-of-domain evaluations.

As tasks require heterogeneous expertise, parallel work, verification, and persistent state, improving a single LLM agent is not enough. Graph Engineering proposes explicit, dynamic graphs for coordinating agents, tasks, and evolving system states.

Instinct is drawing praise in private testing for handling email, travel and other everyday tasks. Its broad permissions, data terms and ability to act without confirmation are also raising serious privacy and security questions.

OpenAI is extending agentic AI beyond software engineering into finance, investing, operations, and other office workflows. Its biggest obstacle may be trust and usability, not model capability alone.

AWS has open-sourced Dogwood, a policy language that can evaluate an agent’s current tool request against its execution history. It extends single-request authorization to workflows and concurrent actions.

Cloudflare’s Agent Tracing adds agent-level observability to Workers, covering model calls, tool execution, approvals and subagents. Payload defaults differ across frameworks, while truncation and upcoming usage fees limit how the feature should be used.

SkillEvo argues that the main bottleneck in evolving agent skills is not merely the ability to edit them or the number of iterations, but whether evaluation keeps producing reliable directions for improvement. It turns multi-turn user simulation into a feedback generator and adds an independent governance layer to control factual and structural degradation.

PolicyGuide compiles organizational policies into workflow graphs and proactively checks agent state at user-turn boundaries. The approach helps customer-service agents recover missing procedural steps instead of merely blocking risky actions.