Agents

This week on The Reg's Kettle podcast, we wonder whether tokenminning is going to bring the industry back down to Earth

Crowdmind is a local-first desktop app for fast qualitative research. It lets you create synthetic AI persona panels and test products, messages, pricing, landing pages, images, PDFs, or multi-step funnels, receiving structured feedback like scores, objections, positive signals, and recurring themes. Supports multiple LLM providers including local offline models. All data stays on your machine in a local SQLite database. Ideal for founders, product marketers, researchers, and product teams.

Outlines is an open-source library that introduces deterministic certainty into LLMs' output generation process for better, more reliable generation of structured outputs.

The article compares six video similarity measurement techniques—GPT Vision, Gemini Flash, CLIP, perceptual hash, CV multi-metric, and Gemini Embedding 2—using a benchmark of waterfall clips. Accuracy is prioritized over speed. Gemini Embedding 2, which processes the full video, emerges as the best balance of accuracy and speed, outperforming frame-sampling methods.

A new wave of social engineering attacks, dubbed 'SociaLLM Engineering,' targets AI agents powered by large language models. These attacks manipulate LLMs into revealing sensitive information or performing unauthorized actions by exploiting their implicit social understanding and lack of trust boundaries. Real-world cases include Instagram account takeovers, GitHub workflow data leaks, and 'Bioshocking' of AI browsers. The article examines why LLMs are particularly vulnerable—due to their design to please users, single-channel processing, and lack of memory—and suggests mitigations such as human oversight and robust guardrails.

Cairn is a self-authoring AI agent operated by Omri Pitaru. It lives in a public GitHub repository where it edits its own personality, memory, goals, and writing. It operates on a fixed budget and communicates via email.

Cloudflare launches Precursor, a client-side behavioral validation engine that continuously collects interaction signals to distinguish humans from bots across full user sessions, reducing friction for legitimate users and improving detection of advanced automation.

OpenAI merged the ChatGPT desktop app with Codex, removing beloved features like screenshot and 'Work with', and replaced it with a Codex-centric interface. The author argues the browser remains the best option for ChatGPT users.

Vairfid provides AI agents with verifiable identity, trust scores, and verification records, enabling secure cross-company collaboration.

Loam is an AI-powered applicant tracking system designed for early-stage founders making their first 10 hires. It combines applicant tracking, AI candidate review, sourcing, chat, and a branded job site into one platform, with simple monthly pricing starting at free. It targets founders who are overwhelmed by spreadsheets or cannot justify enterprise ATS costs.

Cloudflare will block AI agent crawlers by default on ad-supported pages from September 15, categorizing bots into Search, Agent, and Training. This forces AI companies to renegotiate access and spawns pay-per-use models.

DiscoMCP is an open-source tool that transforms any MCP server into a tailored skill for AI agents by analyzing actual usage patterns, rather than listing all tools. It guarantees read-only operation, requires zero setup, and reduces round-trips for complex tasks significantly.

AI-assisted development can quickly generate polished frontend code, but it often misses critical aspects like accessibility, keyboard navigation, focus management, and error handling. The article emphasizes the need for stronger verification practices, including clear engineering expectations, design systems, and behavior-focused testing.

Diff Forge AI is a native, local-first Agentic Development Environment (ADE) that lets you run coding agents like Codex, Claude Code, and OpenCode in parallel, with voice control, screen snipping, and a web dashboard for remote visibility. It features multi-terminal workspaces, Loop Spaces for scheduled automation, cloud sync, and device management. Pricing ranges from free to $2,000/month.

AI Connector is a Magento 2 extension that acts as a unified bridge between your storefront and leading large language models like Claude, ChatGPT, and Gemini, offering a single REST API and PHP integration layer.

Stanford researchers present TRACE, a system that diagnoses missing capabilities from agent failures, synthesizes verifiable training environments for each, trains LoRA adapters via GRPO, and composes them with token-level MoE routing. It achieves +15.3 points on τ²-Bench and 73.2% Pass@1 on SWE-bench Verified.

Meta's Muse Spark 1.1 scores 51 on the Artificial Analysis Intelligence Index, up 8 points from version 1.0 in just three months. Gains are concentrated in Scientific Reasoning, Coding, and Knowledge. The model is token-efficient and cost-effective, with an estimated $0.26 per Intelligence Index task.

This article explores the evolution from traditional to-do lists to intelligent AI agents that automate task management and boost productivity.

Prime Intellect released verifiers 0.2.0, previewing a rewritten v1 core. It decomposes an environment into a taskset (what), a harness (how), and a runtime (where), with an interception server for proxying requests and recording traces. Any taskset works with any compatible harness, with full prime-rl training support.

The article explores how AI amplifies the Dunning-Kruger effect by boosting confidence and splitting actual ability into assisted and intrinsic, preventing the traditional self-correction. It argues that intrinsic skill erosion becomes a governance risk rather than just a productivity loss.

The rapid rise of autonomous AI agents and automation platforms is creating severe hardware bottlenecks, with memory bandwidth becoming a key performance driver. Apple's unified memory, CUDIMM standards, and a new PC upgrade cycle are reshaping the market, while memory manufacturers like Samsung and SK hynix benefit structurally from HBM capacity allocation and limited supply.

BeyondSight is a permanence-aware end-to-end driving framework that decouples actor existence from observability by maintaining persistent actor hypotheses, enabling reasoning under occlusion. Experiments show detection mAP for unobservable actors improves from 0 to 0.249 and planning L2avg reduces from 0.61 to 0.54.

This study identifies vascular metrics associated with navigation difficulty and develops an automated pipeline for quantitative feature extraction to enable future complexity grading. Vascular trees from 61 patients were analyzed using a Soft Actor-Critic RL algorithm for 120 s autonomous navigation. Results show that left-side bovine arch and type II/III aortic arch increase navigation time by 30.19 s and 37.92 s, respectively, while greater tortuosity prolongs procedure and reduces success. On the right side, type II/III arches extend time by 45.94 s, and each additional reverse curve adds 3.96 s. The pipeline provides a foundation for standardized complexity grading and RL model evaluation.

This paper presents Dec-MARVEL, a decentralized budget-aware exploration framework for communication-free multi-UAV teams with directional sensing. Robots coordinate by observing teammate trajectories within their field of view. Using a graph-attention actor, they select return-feasible waypoints. Experiments show superior exploration rates and minimal sensing overlap across various team sizes and budgets, with successful sim-to-real transfer.

SplatCtrl is a unified framework for real-time scene reconstruction and reactive robot motion generation, enabling collision-free robotic arm control in unstructured and dynamic environments. It builds on 3D Gaussian Splatting with hybrid voxel filtering and dynamic Gaussian relocation, derives continuous signed distance functions from isotropic Gaussians, and integrates them into control barrier functions. Experiments validate its effectiveness in simulation, on physical robots, and in shared human-robot workspaces.

AgenticFocus is a Mixed Reality synthesis pipeline that converts ordinary first-person-view human videos into robot-trainable demonstrations by restoring occluded object geometry, reconstructing full-hand motion, and retargeting it to a humanoid embodiment. It achieves lower trajectory error and smoother wrist motion than cross-embodiment baselines, with SPARC scores of -5.18 vs -5.56 and -6.05.

MultiView-Bench is a diagnostic benchmark designed to evaluate vision-language models' ability to integrate observations across multiple viewpoints into a coherent, world-centric 3D mental model. Current VLMs excel at single-view 2D tasks but struggle with 3D spatial relations and cross-view aggregation. The authors propose ViewNavigator, a multi-agent framework that actively selects informative viewpoints and fuses multi-view evidence, achieving 3-5x performance improvements on the benchmark.

Knowledge graphs (KGs) often contain factual errors from automatic construction. AgentKGV proposes an agentic LLM-RAG framework with dynamic routing and iterative query rewriting, enhanced by a two-stage training strategy (distillation-based SFT and trajectory-level GRPO) for improved accuracy and cost efficiency. On the T-REx benchmark, macro-F1 improves by 14.9 percentage points over single-turn RAG, with search calls halved.

KV-PRM is an efficient process reward model that eliminates text re-encoding by directly using the KV cache from LLM generation, reducing scoring cost from O(L²) to O(L). It matches or outperforms text-PRMs on benchmarks with up to 5000x FLOPs reduction, 37x latency reduction, and 34x memory reduction.

The L-MAD framework systematically evaluates multi-agent debate structures and aggregation methods in Legal Textual Entailment. By assigning expert personas, it improves upon single-agent baselines by up to 8%. Increasing agent population reduces inconsistency and improves accuracy, but extending discussion rounds induces over-deliberation drift where agents reinforce each other's mistakes. The findings outline practical boundaries for deploying multi-agent systems in high-stakes legal reasoning.

This paper introduces a neuro-agentic control framework that couples an LLM planner with a time-series foundation model (TimesFM) using counterfactual physics injection to ensure physics-grounded autonomous defense, outperforming LSTM and TCN on SWaT dataset with zero hallucinated actions.

ARCANA is a collaborative multi-agent framework for solving ARC-AGI-2 tasks under strict test-time and hardware constraints. It decomposes each task into iterative perception, hypothesis generation, symbolic execution, and reflective refinement. Using a differentiable blackboard and learned meta-controller, it combines structured program search with adaptive multi-turn correction, improving reasoning efficiency and solution quality on abstract transformation tasks.

Researchers frame the formalization of the Vlasov equation's mean-field derivation as a strategy game, where a mathematician directs an AI system to convert LaTeX documents into Lean 4 proof assistant code. The case study successfully completes a full formalization of well-posedness for the nonlinear Vlasov equation via Dobrushin's mean-field route, including existence, uniqueness, stability estimate, and mean-field limit, as well as a short-time superposition principle. About one-sixth of the formalization yields a self-contained layer reusable by the broader library.

Long-Horizon-Terminal-Bench is a terminal benchmark of 46 long-horizon tasks across nine categories, including experiment reproduction, software engineering, multimodal analysis, interactive games, and scientific computing. It decomposes tasks into fine-grained subtasks to provide dense intermediate rewards and partial credit, offering a more complete evaluation of AI agents. Evaluating 15 frontier models, the strongest achieved a 15.2% pass@1 at a partial-reward threshold of 0.95 and 10.9% at a perfect-reward threshold of 1.0, with mean pass rates of 4.3% and 1.7% respectively, indicating significant room for improvement.

GATS is a new agent planning framework that uses systematic UCB1-based tree search and a layered world model to eliminate LLM calls during planning, achieving 100% success rate. It outperforms LATS and ReAct on synthetic tasks and 12 challenging scenarios with lower computational cost.

AVA is an open-source, self-hosted voice AI agent for Asterisk/FreePBX, offering quick deployment, multi-agent management, real-time dashboard, and support for multiple AI engines. Recent updates include stability fixes, silence watchdog, and per-agent voice selection.

31-year-old voice actor Shen Anyu faces a career crisis due to AI clones of his voice. The clones spread widely, causing platforms to flag his real recordings as synthetic, impacting his income. He and his wife spend significant time tracking infringements, but enforcement is difficult. AI voice cloning tools are disrupting China's short drama, audiobook, and short video industries, with many voice actors facing similar challenges and declining earnings.

Baton is a macOS menu bar utility that monitors AI coding agents like Claude Code and Codex, displaying a live count of sessions waiting for your attention. It uses FSEvents for instant updates and allows click-to-jump to specific sessions.

Clark is a solo-built AI assistant aiming to match Manus agent in features and capabilities. It can use the computer, browser, perform deep research, and integrate with Google tools. Thousands use it daily.

The concept of Directly Responsible Individuals (DRI) originated at Apple and is defined as the person ultimately accountable for a project's success or failure. The author argues that LLM-powered agents should never be considered DRIs because only humans can take accountability. This echoes IBM's 1979 training slide stating that a computer cannot be held accountable and therefore must never make a management decision.

OneDev integrates AI users as virtual teammates that work from issues, create pull requests, review code, and respond to CI/CD failures, keeping all work visible and traceable within the same platform.

Lyzr, a three-year-old Jersey City startup that helps enterprises build AI agents, used its own AI agent SivaClaw to raise a $100 million Series B at a roughly $500 million valuation. The system fielded questions from over 130 investors, drafted investment memos, and tracked which slides backers lingered on, proving the product works.

An Argo CD UI extension that adds an AI-powered assistant tab, allowing users to query Kubernetes resources in natural language with context including manifest, events, and optional logs. Compatible with any OpenAI-compatible backend and requires Argo CD v2.13+.

Recent benchmark results show GPT-5.6 Sol achieves 100% recall and a 0.91 F1 score at $0.70 per PR review, outperforming all Anthropic models. No Anthropic model reaches the frontier; Fable 5 is dominated by cheaper alternatives. Grok 4.5 and Gemini 3.1 Flash Lite offer cost-effective options. The study uses private synthetic repos to prevent contamination.

xysq.ai is a collaborative memory platform for AI-native teams and enterprises. It connects AI tools and apps, captures context from team workflows, builds a living knowledge graph, and provides the right context when agents need it. Features include isolated team vaults, role-based access, document organization, and a strict no-training-on-user-data privacy policy.

Adaptive Recall is a memory system for AI assistants that learns from interactions, using multiple retrieval strategies, cognitive scoring, knowledge graphs, and self-improvement to provide persistent, evolving memory.

Fade Engine is a fully autonomous AI that shorts overextended small caps on a live $10,000 simulated account, posting every trade publicly. It scans 12,000+ tickers every five minutes, identifies 18 pump patterns, and closes all positions by market close. No human intervention.

The article proposes crowdsourcing unused AI inference tokens for scientific research, drawing parallels to SETI@home. It highlights recent successes by small teams using AI to solve math problems and discusses the design challenges of such a platform.

This guide explains loop engineering, where AI agents autonomously iterate toward a goal using a verifier, state, and stop condition. It details Andrej Karpathy's autoresearch loop and Bilevel Autoresearch, showing concrete results: autoresearch found 20 improvements from 700 experiments, cutting GPT-2 training time by 11%; Bilevel Autoresearch added an outer meta-loop for a 5x larger val_bpb drop. It also provides reusable building blocks and a hands-on template.

exxperts is a local-first agentic runtime that provides persistent AI rooms with governed, approval-gated memory. Everything runs locally as files on your disk, ensuring privacy and control. It offers both a web app and a CLI/TUI interface.