All Stories

Factory lets Jira teams assign coding tasks to its Droid agent

A study finds 124 unique tool jobs across 11 coding-agent harnesses

Codex adds a /visualize command to its app

Blume 2.2 adds agent-assisted migrations for documentation sites

OpenCode uses agents to triage issues and open pull requests

Sam Newman argues LLMs lack causal models for reliable operations

llama.cpp ggml RPC backend supports inference across heterogeneous devices

NVIDIA VERA research paper describes 9,000+ restartable agent sandboxes

Sayak Paul explains KV caching for image-generation flow models

Google's AI Edge Foresight transcribes voice notes on macOS

d1 decision models add text, vision and audio variants

Microsoft previews local-or-cloud model routing for Windows Copilot

Nous Research raises $90 million to bring Hermes to enterprises

OpenAI says Codex and ChatGPT Work reached 40 million users

Liquid AI releases open-weight decision models for on-device use

Quail adds query-planned prefix sharing to its AI-SQL engine

vLLM-Omni unifies serving for speech, vision, diffusion, and robot actions

Vercel offers Glyph Cluster for coding and agentic knowledge work

Artificial Analysis adds base-model labels and filters for derived models

PyTorch documents pinned-buffer reuse for CUDA-to-CPU copies on NVLink-C2C

Cloudflare's Clef decision models become available through OpenRouter

Hybrid Latent Attention shrinks KV caches for looped language models

Vercel AI Gateway exposes OpenAI's structured Decisions API

Claude Haiku 5.5's published results show gains over Haiku 4.5

An Xid 31 guide traces production GPU crashes to MMU faults

Prism ML fits Qwen 27B into 6GB while retaining 95% of performance

Datology releases Zephon for deterministic, elastic training data pipelines

vLLM reports 5.3× higher throughput for DeepSeek V4.1 Flash at concurrency

AI networking engineers flag fiber handling as a production bottleneck

ReqLLM 1.27 adds support for the OpenAI Decisions API

tldraw lets canvas comments prompt ChatGPT and Codex agents

LearnOpenCV explains how NMS-free detectors remove a box-cleanup step

Microsoft brings local model routing and file context to Copilot on Windows

NVIDIA's CUDA guide maps kernel and memory controls to AI performance

An 8-million-parameter TTS model runs from just 9 MB of weights

Microsoft plans local MAI-Code-1.1-Flash for routine Copilot coding tasks

Mistral claims Large 4 ranks first among open models on Harvey’s LAB

Microsoft reportedly demos 284B-parameter DeepSeek V4 Flash on a laptop

Posts say Grok Bot will route tasks to outside models, including Claude Opus 5.5

OpenRouter adds Perplexity Decider v1.1 at $0.02 per million input tokens

Google Research presents EnvHarness for adaptable agent training environments

Epoch AI's InnovationEval finds agents struggle to invent post-training methods

Trillium Labs says compute access is blocking open distillation research

A Grok Bot workflow exposes a private agent API tool across an organization

Pi creator publishes subagent extension for adversarial reviews

A Codex workflow routes daily Todoist tasks by explicit owner

Rivet adds Pi Durable agents to Actors with crash-resumable execution

Kernel's Config Registry makes more sites reachable to browser agents

CopilotKit's Manufact cookbook embeds interactive MCP app interfaces in chat

OSWorld introduces an iOS benchmark for agents with persistent user identity

LangChain Managed Deep Agents v0.9 adds schedules and per-run configuration

An eval-review workflow uses diverse agents and privileged trajectory inspection

A first-principles study tests when agent swarms help solve hidden-rule problems

EffectTS's Blacklight harness stores agent state in an append-only log

Salesforce's CLIFT trains web agents to verify rollouts without live judges

DAEDALUS studies agent memory with ablations across its pipeline

Tembo adds the Vercel MCP for deployment and build-log access

Mastra brings Agent Controller, signals, schedules, and goals to GA

OpenHands adds scoped agent profiles and isolated Docker runtimes

Cline launches cloud agents that test code and open GitHub pull requests

LiteLLM team shares Moyai, a self-hosted agent for background coding

ComputeSDK Actions adds support for multiple sandbox providers

Jonas shares an open-source, self-hosted family agent built on Cloudflare

A browser-use test finds Haiku 5.5 close to Opus 5 on defect checks

Hermes background reviews can now improve user-owned skills

ai-rulez generates native assistant configs from one source of truth

Kittens offers self-hosted development machines backed by Firecracker

A team agent routes X requests to the developer who last touched the code

Devin Mobile lets developers manage cloud agents across operating systems

Grok Bot 0.68.1 adds PowerPoint and Google Slides deck creation

An Apple paper reportedly finds a single coding agent beats complex ML harnesses

Deepagents can load tools alongside skills without breaking prompt caching

A coding-agent orchestrator delegates large parallel test matrices to subagents

A LangChain engineer uses an ACP adapter to run Deep Agents in Zed

Geoffrey Huntley says Claude Code’s classifier disrupts his development loops

Zephon open-sources a deterministic dataloader for large-scale ablations

Google Research introduces ContinuousBench for testing private synthetic data

OpenAI rebuilds its plugin submission flow with clearer review feedback

Menso runs AI personas through live websites for usability testing

A production-trace workflow asks agents to find high-impact waterfalls

Devin adds Haiku 5.5 and reports a 58.4% FrontierCode score

OpenAI's 722 math manuscripts shift attention from peak results to research throughput

Ryan Lowe argues safety processes can help teams move faster over time

A Claude Code postmortem traces a C-drive deletion to Windows quoting

Gergely Orosz says nondeterministic agents need deterministic data controls

ClaudeCost estimates $4,853 in API-equivalent cost across 8.6 billion tokens

OpenTunnel launches public URLs for local services with claimed end-to-end encryption

Google opens SynthID Detector globally in English

Mathematicians describe novel techniques in OpenAI's released proofs

A Lean autoformalization paper warns that checked proofs may mistranslate the original

NVIDIA researchers find tool use weakens multimodal models' refusal behavior

Vercel Flags adds timestamp-based targeting for scheduled releases

T3 Code displays an agent's evolving feature graph in its work thread

Doug Colkitt tightens an OpenAI integer-multiplication result

AI agents now account for about half of visits to the FRED data portal

OpenAI paper resolves a 3D elasticity inverse-problem uniqueness question

Anthropic releases Claude Haiku 5.5 at $0.10/$0.50 per million tokens for short prompts

Anthropic releases Haiku 5.5 for faster, lower-cost agent tasks

A coding-agent study finds harness choice changes cost more than accuracy

A developer's Skatebench comparison finds a 55× lower task cost in one year

A developer reports 842 billion tokens of API-equivalent use in September

Anthropic cuts Claude Sonnet 5.5 cache-read price to $0.10 per million tokens

Theo says T3 Code cuts GitHub rate-limit usage by over 75%

Ramp's October vendor data shows cheaper models and media generation pulling AI spend

Chandra OCR creator alleges a project copied Chandra OCR 2

An MIT paper reportedly improves agent memory by exposing history as Python variables

DAIR.AI releases MCP tools for its curated research paper index

Cohere opens private beta for its Compass Cloud retrieval platform

Vik Paruchuri rebuilds chart extraction with fewer errors and faster processing

Turbopuffer makes sharding generally available for namespaces up to 256TB

Unsiloed computes document values that are not printed on the page

Matt Pocock argues coding agents need to care about the codebase

PersonalAgentBench compares assistants on everyday tasks and authority limits

A source-preference study finds LLM agents favor familiar websites

Biohub and partners commit $1.8 billion to data for virtual-cell models

ComfyUI uses Blender blockouts to control AI-video choreography

FrankenManim replaces Manim with a compact Rust animation toolkit

A public index gathers ML training logbooks and postmortems by year

A startup founder says one engineer shipped four polished apps in six months

Huntley publishes Nix overlay walkthrough with a runnable demo

Overmind open-sources training from production agent traces

Intel Panther Lake uses PowerVia backside power delivery

Vercel adds beta fallback models for low-confidence decisions

GPU MODE launches a repository of AI performance lectures

A Devin autoresearch prompt runs autonomous hill-climbing experiments

Model routing should expose the right tools, not guess intelligence

ICML study finds AI and human paper rankings diverge sharply

Gemini Live adds Guided Vision for blind and low-vision users

Mistral Large 4 reaches the top open-weight tier in benchmark testing

Mistral Large 4 preview posts mixed coding and cyber benchmark results

Mistral trains Large 4 in European data centers on 3,800 GPUs

DeepSeek V4.1 Flash scores 72.9% on ARC-AGI-2

vLLM 0.31 adds a daemon to preload model weights

AMI’s H-JEPA raises Visual AntMaze success from 18% to 73%

Mistral Large 4 reportedly used 4,000 GPUs versus Astra's 100,000

Two agents play StarCraft in real time with four tools and no video input

Cortex Harness lets robots hand control to humans when stuck

Google Research demos ScientistTwo for autonomous paper analysis

Harness-aware distillation raises ALFWorld tool-context use to 73.1%

Eleven v4 Turbo ranks first on a text-to-speech leaderboard

Meta MIRA splits long-horizon research between planners and executors

Grok 4.7 posts mixed ARC-AGI-2 and ARC-AGI-3 results

Sesame adds voice control for coding agents and connected apps

Grok Bot users report strong chat but missing computer-use access

Google Research previews Co-Director for coherent long-form video

AGENTS.md workflow removes no-ops and moves instructions to CODING_STANDARDS.md

Google AIM improves scientific-agent search by 1.6 and 4.9 points

OpenAI makes ChatGPT auto-review free for account users

Claude Code 2.1.292 adds effort controls for spawned agents

Parsewave finds 206 faulty graders in 600 AutomationBench tasks

Anthropic expands Cyber Verification Program to offensive testing

Google Drive and Docs add native Markdown files

Gamma 5 launches agentic presentation editing with PowerPoint import

Google Earth AI maps Ebola exposure in the Democratic Republic of Congo

Scrimshaw turns Claude into a playable text-based music generator

Claude edits Google Workspace files directly from open documents

Mistral Large 4 leads a legal-agent benchmark at 15%

fal adds image and video generation to ChatGPT and Codex

Google Earth AI uses geospatial embeddings for public-health forecasting

Codex and T3 Code generate visualizations inside threads

Delta v0.18 adds thread bookmarks with Cmd-D and Ctrl-D

Vals uses 90 Opus agents to find magnetic semiconductor candidates

FrankenTerm cuts headless terminal processing time by 3.2×

NVFP4 outperforms MXFP4 on B200 matrix benchmarks

GPT-6 Astra and Sol reportedly reach 50 tokens per second

AWQ cuts Qwen3-Omni first-audio latency from 213 ms to 23 ms

Together Link routes open models through coding harnesses

ValsAI agents screen magnetic materials with DFT simulations

Z.ai releases GLM-5.3 on Amazon Bedrock

Manus launches an agent-driven video editor

Browserbase introduces cryptographically verified identity for web agents

MCP Events proposal adds asynchronous agent-to-agent work

html-plan plugin turns Claude Code plans into HTML mockups

Cua Driver adds six open-source cursor motions from 82 prototypes

Deepgram Flux TTS preserves acoustic state across turns

Eve launches an iMessage agent stack with memory and multi-tenant accounts

TanStack Start adds JSON-spec app generation with server-side rendering

Report: Figure plans a Hark agent for web and mobile

rho-1 introduces a 19B model for reasoning and robot control

One model plays multiple image-only games in parallel

LiveKit integrates Microsoft AI streaming speech models

AWS guide deploys Pipecat voice agents with Deepgram on SageMaker

Coucou puts coding-agent permissions and sessions in the Mac notch

OpenAI rolls out text watermarking for ChatGPT and Codex in the EU

Zeeg reports coding agents can produce 50 to 60 simple PRs per day

Claude Code 2.1.291 fixes dropped-session responses and final-message loss

SelfSearch says a self-modified harness matched Codex at $4.03

Flow-1 targets lower-cost failure finding in agent traces

New York City proposal would require external validation for AI models

QuiverAI launches an API platform for keys, usage, billing, and logs

WeirdML v3 tests autonomous data and machine-learning work

Qwen3.5-9B paper rewriter opens a public GPU demo

T3 Code adds CSS-themed visualizations inside agent threads

ComfyUI publishes a keyframe animation workflow

Venice adds account-free access to FLUX 3 Image

Amazon reportedly shuts down consumer-agent access to Muse and Instinct

Dot detects a possible account compromise from email and login context

Science Made Readable rewrites open papers without dropping caveats

Uber has fewer microservices in 2026 than it did in 2020

Amp removes Mermaid rendering in favor of ASCII, images, and HTML

A developer uses Opus 5.5 to clean up and test a personal web app

0xblacklight says getting database schemas right early pays off

Robert Martin uses agents to review repositories and report code issues

Practitioner claims an $80 run trained a 4B model to beat GPT-5.6 Terra

Post reports a possible Gemini 4 Argon listing on Google Cloud

Claude Fable 5.5 builds a black-hole renderer in Python

EditHero benchmarks multi-step, part-level 3D editing

BRUT turns a static website into a procedural city-building demo

DMAD distills visual generators without an auxiliary diffusion model

Maxime Rivest argues smaller generalist models should rely on retrieval

Aleph Alpha cites Kolibri scores of 96.9% on AIME and 66.4% on SWE-Bench

Anthropic reportedly adds an opt-in setting for voice-data training

OctLLM represents 3D geometry with explicit octree tokens

Dot user describes daily website digest scheduled through Chrome

Microsoft's Agensh study reports team gains at 128 agents on ProgramBench

Maxime Rivest demos scientific article rewriting attributed to a 4B model

Hermes supports iMessage through BlueBubbles and Photon, Nous says

Ethan Mollick says organizational systems may bottleneck AI gains

GitHarness versions agent work so changed requirements avoid full rework

Jakemor outlines mobile interface patterns for embedded agents

Nick Dobos proposes values-based prompting for proactive agents

A developer adds Bitwarden passwords and passkeys to a personal agent harness

Ralph Loop v2 discussion proposes an external observer for agent runs

GPT-6.1 Sol becomes one developer's default for coding and subagent work

Amp uses screen recordings to specify requests and verify fixes

Dingyi publishes an October 2026 tier list of AI orchestrators

Developer uses ChatGPT computer use to manage eBay hardware listings

DeepSeek Harness tests compatibility with Claude Code Mods

ffmpeg-skill gives coding agents a probe-plan-render-verify workflow

Users report Codex agents feel slower than three months ago

Claude diary entry reportedly leads Anthropic to contact police

FunBench ranks models by human-rated creativity and playfulness

Anthropic recommends Claude Code's /usage breakdown for token spikes

Rocky Ding explains the harness checks behind reliable agent execution

Cockroach Labs' Peter Mattis urges stricter testing for coding agents

VeriHarness releases 26,000 agent rollouts to check shared errors

A developer says autonomous agents got his OpenAI account deactivated

Wired reports Meta Muse builds detailed profiles of users’ contacts

Supermemory presents smfs as a filesystem-style memory interface

Open-source Slack client streams live unread state to Cloudflare Durable Objects

Instinct screens products for birds after misreading a user's joke

Avi Chawla compares five common vector-search index designs

Reducto separates table retrieval from reasoning over table data

LlamaIndex says Extract v2.5 cuts scanned-form errors by more than half

Vercel's Guillermo Rauch separates human and agent documentation

Omar Khattab explores artifact-based human-agent collaboration

A developer animates code-drawn illustrations with Fable 5.5

Rippling's agent queries blur the boundary between tools and data

Decode publishes a free course for building a coding agent from scratch

Dabit urges developers to expand skills and experiment with cloud agents

CodeMarkets offers a CLI to launch revenue for open-source projects

Kanna groups Claude Code and Codex chats into project channels

Matt Pocock's codebase-design skill searches for abstractions to delete

GitButler makes its workspace branch appear only for parallel work

Codex's IDE-style interface wins one developer over terminal workflows

Matt Pocock argues stronger abstractions can constrain coding agents

Yuchenj_UW argues coding agents need an agent-centered desktop interface

Cerebras CEO says its architecture avoids HBM and CoWoS supply bottlenecks

Third Circuit rules Ross Intelligence's Westlaw training use was not fair use

NVIDIA adds a 64GB DGX Spark configuration for local AI workloads

Fireworks launches cache-aware FireRouter for coding requests

vLLM Semantic Router releases Decision 2.0 models from 0.6B to 27B

SGLang 0.5.21 adds live prefill-decode role switching and serving changes

Hugging Face releases huggingface_hub 1.0 with a new CLI and transfer backend

Baseten reports an agent-built Qwen engine beat vLLM on one B200

CAST adapts speculative decoding trees to verification costs

Apple’s LoopCD cuts recurrent compute while preserving model quality

Paperclip users pass one trillion agent tokens in a day

A guide explains why agentic RL depends on reliable environments

RunInfra serves DeepSeek V4.1 Flash at 378 tokens per second

A technical thread argues distillation can weaken data pipeline expertise

ArXiv submissions reach a record 40,363 papers in September

Yacine argues PPO and small recurrent models deserve another look

MagicPath adds an Open in ChatGPT handoff for shared design canvases

Yann LeCun points to robotics and driving as tests of AI's limits

Antigravity CLI runs on Android through Termux

Bend reports 4.2× faster GPU runs, while Mojo closes in on cuBLAS

Shahules786 frames RL environments as saturating training data

OpenAI's Thibault Sottiaux confirms GPT-6.1 Sol Ultrafast

Runinfra launches hosted DeepSeek V4.1 Flash with 1M-token context

Sebastian Raschka publishes a from-scratch GRPO tutorial

Meta research reports post-training can reduce agents' task coverage

Vikram distinguishes sidecar deployment from broader 800VDC adoption

Zach Mueller shares a proxy for running local models inside ChatGPT

A power estimate puts NVIDIA B200 at roughly 4.4 times Huawei 950DT

Franken Code Browser renders a five-million-line project in 3D

Andrew Lampinen explains how reinforcement learning can enable discovery

Report: Meta is developing an Agent Engine called Forge

RankEvolve paper reportedly raises experiment accuracy from 45.8% to 62.5%

Paulius Zt advises using workflows when task steps are predictable

Memory-extraction workflow uses asyncio.gather and Semaphore(5) for bounded fan-out

NVIDIA open-sources OSMO, a Kubernetes orchestrator for robotics workloads

Jev uses a single output token to control voice-agent turn timing

Skill Cutter audits and trims agent skills to their behavioral core

UI Skills packages reusable design guidance for coding agents

OpenClaw documents shared agent sessions with per-person identity

Kunchenguid calls for portable, self-hosted personal agents

Relay lets agents discover and message other agents

Agent Sandbox Protocol separates execution environments from runtimes

Cartesia Sonic 3.6 clones a voice from a short clip in 44 languages

COLM researchers discuss test-time communication for multi-agent systems

Jensen Huang describes an agent harness as an exoskeleton around an LLM

book-to-skill converts technical books into searchable Claude Code skills

CopilotKit shares four self-hostable templates for agent interfaces

Grok Bots give Slack agents shared context through a GitHub repository

Pipecat voice demo uses Jev with about 100ms of routing overhead

DSPy example uses Jev to route prompts between models

MCP reference SDKs patch redirect and task-session authorization flaws

ActiveSaddler applies adaptive curricula to agent harness optimization

A Claude Code Cloud workflow manages long-running work from a phone

A Reachy Mini harness generates expressive robot motion in real time

Claude Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0

T3 Code updates its machine-wide model usage view

Reports: OpenAI safety lead David Robinson quit over deployment risks

Researchers find image-embedded prompt injections affect four clinical VLMs

Harbor evaluation flaw lets agents alter execution traces before scoring

Meta's self-improving harness paper routes tasks across specialized branches

Bug Hunt Bench shows large cost and speed gaps between coding models

Zeeg warns that coding agents can generate low-quality unit tests

A developer reports GitHub suspended an account used by an AI agent

OpenClaw raises its supported Node.js floor to version 24

A Jido developer leaves Sol Ultra running on planned work for ten hours

ScholarEvolve improves agent scores by evolving the harness from research

Robert Martin recommends mutation testing over forcing agents to follow TDD

Robert Martin argues acceptance tests leave code paths unchecked

ActiveSaddler adapts agent-evaluation scenarios as harnesses improve

NerfBench creator says Theo asked him to take down the benchmark

Enrich Arena compares data vendors on hit rate, speed, and price

AgentMail adds inbox pausing and metrics for agent platforms

Robert Martin says he validates agent-written code with automated measurements

Xiaomi releases MiMo-VL-7B vision-language models and an evaluation suite

DART detects multi-turn agent attacks through representation shifts

ColoredCow recounts PolinRider malware rewriting branches across four GitHub organizations

Supabase adds agent tools for investigating production issues and testing fixes

Former OpenAI safety lead warns iteration may not contain future failures

Pangram flags AI imagery across Etsy handmade listings

Simon Willison argues usage services should default to hard budget caps

Pvncher shares ways to reduce image tokens in vision workflows

Browser Use reports GPT-6.1 Sol beats Astra on Bench 2.1 at 7.8× lower cost

Ai2 open-sources AstaBrief for generating cited scientific reports

Supabase raises $150 million and acquires Turso for agent database workloads

Paulius Zt advises pruning agent context for its next decision

Supermemory exposes shared team memory through a public MCP server

OpenAI explains how Dot preserves agent context across Slack and its app

Dot users report a single conversation and missing mobile shortcuts

A Codex memory pattern combines searchable history with temporal rollups

Q2D argues sparse human labels can misjudge strong retrieval models

Walter builds a Doom and Minecraft mashup as a playable prototype

Opus 5.5 improves a browser game with replay and performance fixes

Fable 5.5 turns a question about LLM training into an edited video

A16Z chart shows paid AI subscriptions in 2.2% of US households

Astra and Opus 5.5 turn thousands of Japan cycling clips into a film

An agent workflow uses Suno tracks and lookbooks to guide video edits

Hamel Husain teaches candidates to handle AI eval interview tasks

Plannotator lets developers answer agent questions asynchronously

Intercom describes standardizing engineering work around Claude Code

OpenClaw Android update goes live after a review delay

OpenHands reports SOC 2 Type II compliance for agent deployments

Uber's existing developer platforms provide verification paths for AI coding

Fable turns one prompt into a scripted and scored animated film

Robert Martin releases crapper, dryer, and mutator tools for four languages

Vercel adds Jev support to its AI SDK for Python

JetBrains opens Air early access for running multiple coding agents in IDEs

Supabase puts schema changes and backend deployment into coding-agent workflows

Claude Code adds JavaScript and TypeScript mods for session customization

Nathan Lambert launches Trillium Labs for open post-training recipes

Google Research introduces TEE-based federated learning for verifiable differential privacy

Perplexity reports pplx-decider-v1-27b scores 85.7% across 11 benchmarks

Research writeup claims off-policy data causes SFT's generalization gap

Venice reports 267 September updates, including agent chat for all users

Grok 4.7 reportedly rolls out across four Grok app modes

User reports apparent Fable 5.5 access in Claude's web app and Claude Code

SemiAnalysis estimates Feynman could cut memory-controller die share to about 4%

DeepSeek V4.1 Flash redesigns its architecture to reduce context-cache reads

Google Research shifts federated-learning computation server-side with verifiable privacy

JERA, Dell and RHAELM plan a 400MW AI data center in Japan

ISTA-DASLab's Qwen quantizations attract millions of Hugging Face downloads

Google automates NCCL ConnectX setup for its GPU clusters

Keras outlines pluggable MLX and PaddlePaddle backends with shared op tests

Sharpening Tax measures lost test-time scalability after post-training

llama.cpp adds `/v1/systemone` for local decision-model inference

OpenEnv adds a capture proxy for RL training in unmodified coding harnesses

KV-streams removes discarded turns from the live cache, claiming nearly 2× speed

A developer builds a personal voice agent with Telnyx, GPT-Live and Cloudflare

Vercel AI SDK fixes an ACP harness flaw allowing unrequested host-tool calls

BeyondTrust details a Vertex AI Agent Engine sandbox response-channel hijack

GitHub previews code-defined dynamic workflows across Copilot CLI and its app

Apple plans additional macOS Full Disk Access controls for AI agents

Stas Bekman shares an agent skill for auditing compute clusters

A developer builds a personal voice agent with Telnyx, GPT-Live and Cloudflare

Kunchenguid maps skills, MCP, harnesses, and models to a software stack

Claude Code CLI 2.1.288 adds recovery for drafts cleared with Ctrl+C

Inworld acquires Ultravox to expand its real-time voice-agent platform

Matt Pocock outlines a GitHub Actions workflow for coding-agent factories

TestingCatalog reports hidden Gemini Desktop settings allow broad file and network access

Claude Code's next-steps plugin helps manage parallel sessions

Reka releases Apache 2.0 RIDM weights for extracting actions from video

GitLab patches a critical Duo AI Gateway sandbox escape

Apple Research questions whether coding agents need complex orchestration

Researchers identify a ChatGPT macOS flaw that could expose local data

Modal adds VM sandboxes and sidecars for agent execution

Anthropic prevents new Claude shared chats from appearing in search

A review of 1,123 live Bolt apps finds exposed keys and open databases

A study finds reasoning models can automate multi-turn jailbreaks

AgentBreaker tests context-aware indirect prompt injection in web agents

CloudSEK traces MALFEX malware to compromised npm packages

Meta's SWE-sweep tests whether agents can find bugs without issue hints

Hamel Husain recommends binary checks over 1-to-5 LLM eval ratings

Vik Paruchuri shows five ways PDFs misrepresent text positions

A documentation folder rename lifts an agent's evaluation score by five points

Factory adds model-level consumption analytics for coding agents

Google's Cogentic uses separate agents to explore and verify proofs

A malicious dedh-devops-automation package steals AWS credentials

AgentXploit rediscovers known agent-framework flaws at 59.3% success

OpenClaw v2026.9.8 adds GPT-6.1 Sol and fixes agent reply handling

OpenAI DevDay notes highlight four levers for agent cost control

NVIDIA reports 62.8% lower accuracy on 128K-token tasks than 4K-token tasks

LlamaIndex reports 93–96%+ accuracy for Extract v2.5 on long-list extraction

Researcher reports deciphering 1593 letters with GPT-6 Astra

SvelteKit 3 ships a migration command for its breaking changes

Charm's Crush terminal streams full ANSI command output in real time

Peter Steinberger says an Apple developer agreement blocked synchronized releases

Etalon evaluates streamed inference with token deadlines and fluidity

Fastino releases GLiDE with a Decision Index score of 64.81

Google and Planet launch Project Suncatcher prototype with four Trillium TPUs

Sara Hooker announces a technical report on creating training data from scratch

Tinker says Fulcrum targets author-specific writing through model training

Ai2 releases Olmo-core 3 for scalable mixture-of-experts training

AWS open-sources Strands Decider 2B for agent decisions

Tencent leases 100,000 AI chips from Oracle for five years

xAI leases underused Colossus GPU capacity to Anthropic and Google

InterPositive fine-tunes private video models on each film's production footage

Broadcom may lend Anthropic up to $42 billion for chip leases

Nemotron's MOPD results show teacher models need compatible training pipelines

Every's open-model guide recommends starting with small task-specific trials

fal exposes Grok Imagine Video 1.5 Lite through text and image APIs

Quail estimates a theoretical 6.6-second AI filtering bound on one H100

Cursor adds GLM 5.3 and GLM 5.3 Flash models

Users report web queries apparently routing to Fable 5.5

Broadcom may lend Anthropic up to $42 billion for chip leases

Microsoft launches MAI streaming transcription and two voice models

CoreWeave opens private preview of serverless GPU sandboxes

Meta's Muse joins Tailscale networks as its own node

Google releases RRSI to improve agent harnesses around frozen models

Pi-GUI previews extension-defined desktop views and cards

Vals AI tests Gemini 4 Argon on Kerbal Space Program return missions

Browser Use reports GPT-6.1 Sol's top Bench 2.1 score at lower cost

Tavus says Griffin fooled 48% of participants in live video calls

Meta researchers report a controller raises ProgramBench score from 63.7% to 71.5%

OpenAI launches Dots agents for ChatGPT Pro

Alibaba releases Qwen-Audio-3.1-Realtime for duplex voice agents

GitHub opens Copilot computer use for desktop apps in public preview

MiniMax open-sources OpenAgentCore as a self-managed Agents API backend

Pi GUI lets extensions build custom views for agent sessions

OpenAI Astra models improvise live drums and bass over piano

Kun Cheng releases a Grok bot backed by his engineering knowledge base

Arena runs up to 600,000 isolated E2B sandboxes per day

DevPulse turns repository metrics into human-approved improvement plans

Photon raises $4.5 million to help developers deploy agents in messaging apps

An autonomous AI attacker breaches Dutch vulnerability group DIVD

Sonnet 5.5 tops Artificial Analysis's Coding Agent Index with a score of 68

Android Bench 2.0 adds roughly 30 long-horizon development tasks

Artificial Analysis adds refusal rates and fallback models to its coding-agent index

A Codex command-injection flaw exposed GitHub tokens through branch names

DirtyBlanket packages spread a Linux worm through fake Express modules

CoT monitoring research highlights a fragile AI safety signal

An agent builds a portable CPU benchmark suite for $114

CoreWeave Agent Lens surfaces recurring issues across agent conversations

Study reports agents earned 27.3% of human returns on year-long decisions

Effect 4 cuts bundle size and memory while raising task throughput

Sentry's toolkit connects coding agents to errors, traces, and issue triage

EvalEval and UK AISI standardize reproducible benchmark cards

Epoch ranks Claude Opus 5.5 first on its Capabilities Index

ARC Prize publishes milestone winners' ARC-AGI-3 solutions

Claude Sonnet 5.5 classifiers can move a chat to Sonnet 5

Ofir Press says new coding-agent benchmarks remain difficult to find

Vals AI argues model benchmarks should report cost per task

OpenAI says GPT-6.1 Sol has returned to expected speeds

User reports Astra Ultrafast exhausted a $500 weekly allowance in 15 minutes

Cline says DeepSeek V4 Pro cut two coding tasks from $8,500 to $300

Unverified post claims Gemini 4 Argon supports one million output tokens

Datalab shows Chandra OCR transcribing handwritten technical documents

Comfy Agent helps create an animated short in three focused days

Claude helps a developer build an interactive game animation editor

Karpathy recommends ASD-STE100 and diagrams for clearer model outputs

Users share animation and game demos they attribute to Fable 5.5