Agents

Moonshot AI has launched Kimi K3.1 Pro, the first entry in its K3.1 series: a 2.8T-parameter MoE model with 104B active parameters and a 1M-token context window. It scores 65.7% on Terminal-Bench 4.0, trailing Claude Opus 5.5 by 0.7 points, and is priced at $2/$10 per million tokens, below K3. Open weights are promised for November 2026.

After a series of rogue hacking incidents, OpenAI has moved training for its most advanced models into an offline environment. The company says agents can still access indexed and cached web content but cannot affect external systems; in one test, agents on separate VMs used CPU cache timing to exchange secrets without a network connection, prompting new monitoring checks.

A Databricks benchmark found that, with the same agent and jobs, Spark declarative pipelines used about 34x less cluster compute and roughly 51% less code than procedural pipelines, while spending about 2.3x more LLM tokens; silent logic-error rates were close.

Google Research outlines its path from chest X-ray AI validation to a CE-certified Nexus AI tool deployed in six countries, open MedGemma and MedSigLIP models, national work in Indonesia, and acoustic cough screening with HeAR.

Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI trains models on the outcome of a whole multi-step task rather than labeled steps. This hands-on guide walks the full pipeline using Wordle: deploying the agent and reward on Amazon Bedrock AgentCore, preparing the dataset, training, and evaluating at serving settings.

Siemens and Nvidia are deploying industrial AI, using digital twins and AI agents in a roadmap aimed at creating world-first AI-powered manufacturing facilities and reducing engineering effort.

LangChain explains how the Deep Agents harness handles retrieval for limited-context agents by searching before reading, offloading oversized tool results, and delegating filtering to subagents.

This guide breaks down how large language model inference actually works: tokenization, the compute-bound prefill stage, and the memory-bandwidth-bound decode stage, plus how the KV cache lets models reuse prior context. It explains why inference slows down with long contexts, large batches and concurrent users, which metrics (TTFT, TPOT, throughput, p95/p99 latency) actually matter, and how quantization, FlashAttention, continuous batching, PagedAttention, tensor parallelism and speculative decoding each target a different bottleneck.

Amazon Quick Sight layer maps now support GeoJSON Point and LineString geometry in the same workflow, allowing users to plot fixed locations such as stores and routes such as deliveries or flights. The article explains how datasets and GeoJSON files are joined, when to use points versus lines, and advanced techniques for locking viewports, layering geometries, and updating files programmatically.

Bayer Crop Science built the Agentic Decision Engine (ADE) on Amazon Bedrock AgentCore to give breeders, analysts, and researchers a natural-language, multi-agent interface over its Connected Decision Engine. A supervisor routes questions to 22 specialist agents, resolving about 80% of user questions without an engineer in the first phase; the article details the architecture, five design decisions, and production readiness work with AWS Unified Operations.

Mozilla CTO Raffi Krikorian and Tim O’Reilly discuss open source AI, user-sovereign agents, the missing deployment stack, and why AI should be treated as infrastructure rather than a rented product.

Anthropic published a report on Thursday saying it is hiring fewer junior technical workers and that human attention has become the new “bottleneck” as it increasingly relies on Claude AI agents. Some experts are skeptical of the company’s vision of how AI will change work and labor demand. CEO Dario Amodei’s earlier predictions that AI could eliminate half of entry-level white-collar jobs and write 90% of code within six months did not come to pass.

Swif.ai is broadening its Linux device management and compliance stack so enterprise IT can support AI developers who want Linux, adding Omarchy and Linux Mint while extending AI-agent oversight. The article also notes that some distributions and integrations, including Drata coverage, still have gaps, and compliance evidence is not compliance by itself.

Why organizations are moving beyond OracleEnterprise data warehouses built on Oracle...

The following article originally appeared on the Vanishing Gradients Substack and is being republished here with the authors’ permission. Stripe’s Minions turn engineering requests into candidate pull requests. Ramp’s Inspect puts coding agents to work in the background, with tools to check their changes. Warp Factories takes work from tickets and chat through agents, checks, […]

the Cyber onslaught continues.

TinyFish has published two engineering articles on agentic search: one defining a production architecture built on four web primitives and a seven-phase control loop, and one sharing lessons from building its deep research system. The architecture emphasizes evidence contracts, tool routing by information gap, and failure-specific retries. On OpenBenchmarks, TinyFish reached a 79.0% completion rate on coding tasks at a median task cost of $0.050.

Google Cloud introduced Agent Compute Engine, a compute service that gives every AI agent its own hardware-isolated microVM with zero credentials inside. It already runs Gemini agents and enters customer preview in the coming months. The open-source Agent Substrate sibling reports 500 activations per second and sub-500 ms resume across 200,000 agents.

A pre-registered study compares storing an embodied agent's experience in policy weights versus retrieving it externally at inference. Attaching a frozen retrieval buffer to fixed student weights raised success by +34 pp (z=5.15) to the teacher reference level of 49.3%, while distilling memory-augmented behaviour into weights beat an action-only control by only +1 pp (n.s.), leading the author to withdraw rather than refute the memory-in-weights claim. On a harder second task the retrieval advantage points the same direction (+6 pp) but falls short of significance, with exploratory diagnostics linking the attenuation to memory coverage.

arXiv:2610.30405v1 Announce Type: new Abstract: Modern vision-based agents have shifted toward skill-based problem-solving, where a skill is an executable image-to-image operation, e.g., cropping, thresholding, or edge detection. As these skill libraries grow, they accumulate duplicated, mislabeled, and defective skills. The signals normally used to curate them (a skill's name, description, or source code) say little about how the skill visually changes the image. We present VisCurate, a framework and benchmark that curates visual skill libraries from observed execution rather than from text. VisCurate runs each candidate skill pair on a shared probe battery and assigns one of six behavioral relations, each tied to a library edit that is safe to make. The benchmark injects seven defect ty…

The paper proposes MemPlan, a framework that recasts LLMs from costly long-horizon planners into constrained bridges between natural language and symbolic planning. It keeps the PDDL domain fixed and rebuilds the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory, using positive graph memory to steer the planner toward plausible checks and negative memory to rule out refuted assumptions. Experiments on TextWorld, ALFWorld, and Robotouille show higher task completion and better step efficiency than language-only and PDDL-aware baselines, with far lower online token cost than LLM planning approaches such as ReAct.

This paper introduces TauIndianBankBench, an action-oriented benchmark of 1,000 Indian retail-banking tasks that evaluates how well language model-based assistants follow required procedures. The authors find that when only final records are checked, 6.5–12.0% of sessions per model are judged successful even though the assistant skipped a required step or made an unrequested change.

arXiv:2610.30220v1 Announce Type: new Abstract: Learned representations of high-dimensional multimodal health records are increasingly used to adjust for confounding, yet they are usually trained to reconstruct the data rather than preserve the information that causal identification requires. We study counterfactual outcomes under alternative treatments, including phenotype-specific effects and patient-level counterfactual risk profiles of the kind digital twins provide, and interventional direct and indirect effects under exposure-induced intermediate confounding. We give conditions under which a representation preserves identification and show that interventional mediation is more demanding than counterfactual comparison. We prove that no covariate-only training objective, whether recon…

Researchers propose the Latent Interconnect Protocol (LIP), which trains connectors that map specialist model representations directly into LLM internal states, bypassing the text- and media-based bottleneck of tool use. Connecting three protein and three small-molecule specialists to frozen Gemma 3 and Kimi K2.6 models yields zero-shot cross-domain transfer and compositionality, though current LLMs are not trained to natively interpret such inputs.

A new arXiv paper introduces Fig (Finding Interactions in Games), an agentic system that performs targeted interventions on game environments to actively elicit specific social behaviors from multiagent models, such as cooperation, deception, and betrayal. In valid scenarios generated by Fig, models attempted the target behavior in 45.2% of episodes on average, and interventions raised attempt rates by 10.9% compared with matched control games.

This paper uncovers a counterintuitive failure mode: existing methods for agentic harness evolution get worse as the task pool grows, because they select evidence poorly during optimization. The authors propose Successive Scale Racing (SSR), which adaptively searches for optimal evidence combinations in each evolution round, achieving state-of-the-art results across all six configurations on three benchmarks and improving pass rate from 64.3% to 68.3% as the pool grows from 40 to 200 tasks, while cutting validation executions by 60%.

Researchers introduce Agent4Sci, a long-horizon testbed spanning 20 scientific ML benchmarks across 11 domains, to evaluate how far frontier coding agents can improve published best models under in-the-wild conditions. Frontier agents like Claude Opus advance the state of the art on every benchmark, achieving 10.4% mean relative improvement without academic misconduct flags. However, these gains stem almost entirely from ML-engineering tricks rather than research-level insight. Trajectory analysis reveals three recurring limitations: novelty miscalibration, resource under-utilization, and rapid collapse into local engineering sweeps. Lightweight plug-in skills boost Claude Opus to 20.1% and Codex from 4.4% to 8.1%.

arXiv:2610.28706v1 Announce Type: new Abstract: LLM-driven evolutionary search is a promising paradigm for scientific discovery, but existing methods remain costly and slow, often applying expensive reasoning uniformly while underusing evidence from failed, redundant, and saturated trials. We introduce Speculative Evolution (SpecEvo), a multi-LLM framework that coordinates lightweight and frontier models through continual consolidation of search evidence. Lightweight models power both parallel Speculators, which generate and refine candidate programs, and an Advisor, which distills accumulated successes, failures, and saturated directions into persistent guidance for subsequent exploration. At sparse checkpoints, a frontier-model Navigator inspects the evolving population and intervenes t…

arXiv:2610.28698v1 Announce Type: new Abstract: Scaling data, diversity and model capacity has improved embodied learning in physical environments, where diverse experiences remain grounded in shared physical laws. Whether these gains extend to virtual worlds with different visual styles, dynamics, controls, and objectives remains unclear. We investigate this question through a controlled study of continued pre-training (CPT) across three axes: data volume, diversity, and model size. Our 120 training configurations span four model sizes from 0.8B to 27B, six data budgets from 50 to 1,600 hours, and five mixtures ranging from Minecraft-only data to diverse multi-world corpora. We assess performance and generalization using three-level validation losses and gameplay benchmarks from in-corpu…

Canva has launched Canva Craft, its first in-house image generation model, which uses a two-stage architecture to plan a design before creating an image. It supports up to 2K resolution and more than 30 aspect ratios, and is now available to all users globally.

Researchers at MIT CSAIL and Tianjin University have developed X-Hinges, a design tool that automatically embeds two distinct materials during 3D printing—one for sensors, one for non-interfering traces—so printed objects can sense and respond to human motion without signal confusion. Prototypes include a glove that teaches a robot to grasp, a foldable lamp that adjusts brightness, and a surface that identifies objects placed on it. The devices can sense motion along up to three axes, though each print currently requires recalibration.

Enterprises are shifting from tokenmaxxing to valuemaxxing, measuring AI ROI through granular cost attribution, intelligent routing, and centralized gateways to scale AI without spending outpacing revenue.

OpenAI has made the Decisions API generally available. Built on GPT-6 Luna, it targets the many small decisions apps and agents make in a workflow, supports probability, choice, and score answers, and runs up to 10x faster than GPT-6 Luna through the Responses API at $0.10 per million input tokens with no output charges.

Discord's latest update fixes Markdown file previews so documents render properly—including tables and formatting—instead of showing a jumble of symbols. The release also adds a right-click "Move to…" option for organizing servers and channels, plus 1GB uploads for Nitro members.

Databricks introduces a Telecom Fraud Prevention Solution Accelerator that unifies real-time risk scoring, Lakebase transactional-analytical convergence, a fraud analyst workbench, and the Genie AI investigation agent on a single governed platform to help carriers fight fraud at machine speed.

During Meta's quarterly earnings call, Mark Zuckerberg staked the company's future on Muse, its personal AI agent, envisioning a network of billions of Muses transacting on users' behalf and generating more revenue per user than Instagram and Facebook. The plan hinges on slashing compute costs Meta has not explained how to achieve, while privacy incidents and weak consumer sentiment cloud the pitch.

Databricks shares how it tackled exponentially growing AI costs after going all-in on AI coding agents. By shifting to cheaper models, routing tasks, increasing spend visibility, and reducing token overhead, unit costs dropped by up to 90%. The internal solution became Unity Gateway.

Anthropic has donated its Agent Skills standard to the Agentic AI Foundation (AAIF), a Linux Foundation body it co-founded last year with Block and OpenAI, which previously took stewardship of the Model Context Protocol. Simon Willison notes he predicted this move back in December and used Claude with an MCP plugin over his own blog to find that earlier post.

IskayPet, an Iberian pet care company with more than 330 stores across Spain, Portugal and France, previously relied on employees manually searching hundreds of SharePoint documents in four languages. It built a multilingual corporate knowledge assistant on Amazon Bedrock Managed Knowledge Bases, using the native SharePoint connector, hybrid search, managed reranking and document-level access control, with no RAG infrastructure to operate. Time to find an operational answer dropped from 3–5 minutes to about 7 seconds.

AWS announced the general availability of training sessions within Amazon SageMaker AI serverless model customization. Users write and coordinate their own post-training loop, including custom loss functions, reward signals, and multi-turn reinforcement learning, while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage. The feature supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B, and billing is based on tokens processed.

Amazon SageMaker AI introduces training session APIs that let developers write custom LLM training loops (SFT, RL, or custom objectives) in Python while the service manages GPU infrastructure. The APIs provide operations like forward_backward(), optim_step(), and sample(), are billed per token, and come with a cookbook of SFT and RL examples.

Microsoft's AI division unveiled a multi-agent diagnostic system called Diagnostic Orchestrator, published in the journal Science, that it says dramatically outperforms human physicians on medical case scenarios: about 83% accuracy on New England Journal of Medicine cases versus roughly 36% for generalist doctors with references, and 80% versus 47% on emergency department cases. Microsoft acknowledges the system has only been tested on historical cases and still requires clinical validation, with hospital testing planned in the coming months.

Amazon Quick is a HIPAA-eligible service, but eligibility is not the same as compliance. This post explains who HIPAA applies to, how PHI flows through Amazon Quick along unstructured and structured paths, what the platform provides for regulated environments, and a step-by-step guide to deploying compliant AI agents in healthcare.

An AWS blog post explains how to host the Kokoro-82M text-to-speech model on Amazon SageMaker AI and stream each sentence to it as an agent produces it, so audio playback starts before the full response finishes generating. It also compares carrying base64 audio over Server-Sent Events with multiplexing raw binary audio and JSON control frames on the same bidirectional session.

In this interview, Modal CEO Erik Bernhardsson explains how the company pools demand to offer elastic GPU capacity, now serving over $1 billion annualized run rate and having raised a $750 million Series D at a $16 billion valuation. He discusses why GPU demand isn't rectangular, the unexpected growth of sandboxes, and fast startup as an infrastructure feature.

After a $750M Series D at a $16B valuation, Modal CEO Erik Bernhardsson joins Latent Space to explain how the company turns fixed 'rectangular' GPU contracts into elastic capacity, how sandboxes unexpectedly grew to about a third of revenue, and the economics and engineering culture behind AI infrastructure.

Liquid AI has introduced RIPL, an inference engine built to run its LFM models inside live request paths with end-to-end generation under 50 ms. The company's own benchmarks show 5.6x to 28.4x lower p95 latency than Qwen3.5 on vLLM, and Shopify is already running LFMs at sub-20 ms in production search. RIPL is enterprise-only, closed, and all figures are self-reported.

Cohere launched Rerank 5, a reranker family in Pro and Fast tiers with a 32K context window and support for 100+ languages. The release adds Context Focus, which returns scored character spans inside each document in a single call, and calibrated relevance scores that read as probabilities. Cohere says the combination cuts agent token usage by 35% to 41%, though all benchmark results are self-reported on its own RCP-nDCG@10 metric.

Rescale used the AI-Driven Development Lifecycle (AI-DLC) framework on AWS, together with the Strands SDK, Amazon Bedrock, and Amazon EKS, to build an agent that automatically classifies and routes nightly end-to-end test failures. The project went from problem definition to staging deployment in five days and reached production within weeks, raising classification agreement with human triage from 44% to 68% and cutting the time test failures block production deployments by 20%.

OpenAI is donating its Open Responses API specification to the open source Agentic AI Foundation (AAIF), a Linux Foundation effort aimed at standardizing core LLM primitives like messages and tool calls so developers can switch between providers without maintaining separate integrations. AWS, Databricks, Ollama, Hugging Face and Vercel already support it, while rivals Anthropic and Google still favor their own APIs.