All Stories

Thdxr says AI makes moving from a monorepo to multiple repos cheaper

Matt Pocock proposes Redux-style state control for agents

Omar Sar describes applications that evolve through agent-driven changes

RepoPrompt CE 1.8.0 improves large workspaces and session recovery

Badlogicgames starts Pico, a language designed for codemode

Thdxr switches from monorepos to multiple repos over supply-chain risk

Ray Fernando recommends testing folded layouts in the iPhone Duo simulator

Nix offers a single-source approach to patching software fleets

Geoffrey Huntley argues AI-generated code does not remove the need to fund libraries

A walkthrough shows agents restoring and maintaining old repositories

FrankenGit proposes an immutable Rust forge for agent-authored changes

Superwall embeds native paywall components directly into the DOM

T3 Code adds managed policies for organization-wide configuration

Claude Code adds mid-turn ultrathink to raise reasoning effort

Ivan Leomk runs Cloud Run with SQLite stored on GCS through FUSE

Pele renders Mermaid diagrams to SVG without the Mermaid.js runtime

A Solana developer publishes ARM64 Agave binaries through GitHub Actions

Paolo Anzani says keeping agent code in fewer files can improve output

Alibaba engineer flags architectural erosion from AI-generated quick fixes

Nous Research says beta enrollment has reached platform caps

Devin reportedly ports a model to Trainium in 36 hours

VisMatch v1.4.0 adds three image-matching algorithms

VisMatch unifies access to more than 50 image-matching models

Mistgate offers a low-footprint, self-hosted VPN management panel

Together describes keeping post-training rewards aligned with serving

A researcher frames cheaper distilled models as a step toward AI-driven R&D

Badlogic argues sample efficiency is the real frontier in AI research

Doodlestein recommends Qwen as an open-weight TTS option

Yacine questions how much of the deep-learning stack remains bug-prone

A YOLO11s color-channel swap cuts COCO128 recall by 10.5 points

Jun Song says a group offered money to train models on exploit data

A tester posts code for demos attributed to Fable 5.5

Hacker Papers surfaces arXiv papers through Hacker News discussion

AutoResearch experiments show a training trick can overfit after setting changes

A DIY smart radiator setup uses sensors to model room heat flow

NVIDIA AgentX reportedly tops 1,000 tokens per second per user

MiniMax details local quantization options for M3.1 Flash

A practitioner uses Argon to compare GPT-J training-loss ablations

LookAtMyQuant allocates quantization formats across model components

AI startups resell rented GPU inference while competing on price

Devin describes always-on agents monitoring chat channels

Browser Use says Folk’s Pi harness switch cuts token use

Plannotator releases an experimental inbox for tracking agent work

Yann LeCun argues task performance does not imply human-like intelligence

Wispr Flow notes become accessible to assistants through MCP

Hermes adds Grok Bot and OpenBot import commands

Paper’s developer documents a Windows and WSL MCP connection workaround

Paper recommends mirrored networking for MCP connectivity from WSL

A 2048 demo clarifies that Sol 6.1 used code to choose later moves

A developer spends millions of tokens on repeated agent review rounds

Ofir Press questions how agents can plan around user needs

Grok Bot previews an active-task view and messaging controls

ModelScope expert offers three questions to ask before delegating to AI

Antigravity reportedly tests a coordinating agent called Chief of Stuff

Developer ports Device Manager to Android phones

OpenBot offers a self-hosted AI coworker with separate agent workspaces

Will Depue questions how general models may change robotics ML

A developer builds an allowlisted agent-plugin SDK workflow on GCP

A developer argues custom harnesses preserve control over agent capabilities

Omnigent adds MCP connection checks and pull-request management

Matt Lam compares Codex, Pi Durable, and OpenCode code modes

Uber CAPRI captures production traffic for on-demand replay

DeepSeek V4.1 Flash 1017 reportedly matches GPT 5.5 on an index

Terence Tao warns automated proof-finding can strip-mine mathematics

DevHunt tracks new CVEs and sends stack-specific security alerts

FrankenSim developer explains why open TTS struggled with long narration

Seedance 2.5 prompt coordinates reference images, camera motion, and continuity

WATCHA talk warns agents may bypass tests instead of fixing bugs

Meta and NYU paper reports h1 solves 598 PutnamBench problems

Nous says it blocks providers training on paid OpenRouter usage

Hermes 0.22.6 will fix missing Discord results after name changes

Observers find hidden Mythos 5.5 references in Anthropic website source

Gemini 4 Carbon reportedly tested on Arena as Argon 4

destructive_command_guard blocks risky commands before agents run them

Alibaba engineer recommends pairing AI coding speed with human rigor

Antigravity session log shows repeated model-provider failures

Vercel completes Linux kernel updates for CVE-2026-98432

Observer disputes alleged Anthropic Numbat leak, citing random-number matches

Pi Schema compares JavaScript validators on speed, memory, and conformance

A Rust developer points to memory safety across a large codebase

BLA hardens shipped builds against AI reverse-engineering agents

WATCHA speaker shows how Codex fixes can reintroduce regression bugs

Gemini 4 models post strong scores across research and coding tests

LlamaParse adds dynamic reasoning to Extract Max Effort

A developer replaces 3D models with generated sprites in a fighting game

Amp adds iOS widgets for following coding tasks from the home screen

Hermes plans web access to its desktop interface

Paper founder says pricing will change after community feedback

Vatsal Trivedy routes game-design references to a multi-model agent swarm

Gergely Orosz highlights a game team’s 1,000 pull requests in four days

omux runs Claude Code and Codex sessions in a native Mac terminal

Webpoke proposes consumer-owned cursors for reliable webhook processing

T3 Code adds /btw side chats across models and harnesses

fofrAI suggests turning agent-built sites into installable PWAs

Buck2 cuts a project build from 28 minutes to 78 seconds

OpenAI says it will honor a shared robots.txt signal for AI training

Addy Osmani builds playable Trick or Clawd demo with Opus

Meta's PAMO study reports 47.9% top-10 hit rate with cross-app history

Gemini Carbon generates a Three.js animation in a developer demo

A local Blender workflow combines generated code with existing 3D assets

OpenAI seeks deeper Samsung ties as AI memory shortages persist

Anthropic’s risk factors emphasize compute over customer demand

Baseten reportedly triples annualized revenue to $600 million

Theo says Opus 5.5 high effort beats extra high for most coding work

An upcoming Claude model builds a voxel kingdom from one prompt

A developer runs a Pokémon game workflow with local Qwen models

H-Net learns to segment multi-byte Chinese characters during training

Together AI plans a 250 MW India data center for inference

Batch-invariant mode keeps inference outputs stable across batch sizes

Decision-model tutorial builds a calibrated Qwen3-0.6B example

Gergely Orosz challenges complaints about developers buying RAM for local inference

Explainer traces inference variability to GPU batch composition

Braintrust compares five decision models on 2,000 aerial images

Stageshow gives coding agents a repeatable workflow for animated UI demos

Browser Use opens BU Ultrafast to the public

Practitioner compares TP2 and TP4 DeepSeek V4.1 Flash 1017 layouts across four Sparks

Tembo launches sandbox VMs starting at $0.11 per hour

Practitioners compare low-effort critique loops with multiple-critic workflows

Theo’s computer-use benchmark finds large differences across operating systems

CopilotKit adds configurable layouts for agent tool calls

Google says Gemini agents helped build a robot from scratch

DeepLearning.AI highlights breaches caused by weak agent test setups

Vatsal Trivedy lists harness components that shape agent products

DeepSeek publishes a prototype for repository-level autonomous work

Mikhail Parakhin recommends four critics in a critique loop

A swarm experiment finds public channels aid agent coordination

OpenAI's Astra reportedly beats specialized models on a gripper task

Browser Use reports 22-second median tasks at $0.006 each

Pi Durable raises 4,096-conversation throughput from 46 to 772 runs per second

OmO's Oh My OpenAgent repository passes 70,000 GitHub stars

Cloudflare Durable Objects can coordinate rotating OAuth tokens for agents

ChatGPT desktop voice users report pauses and missed requests

LangChain's Harrison Chase says agent engineering challenges keep changing

Devin tunes an ML workload to match production GPU accuracy

LangChain's Vtrivedy describes bots provisioning infrastructure for research

Report: Antigravity may add local-to-cloud task switching next week

Composio introduces a firewall that checks agent tool calls and responses

Agents may recognize creative signals shared by other models

PokeBot launches RoboAPI for camera-guided robot-arm control

Asmov-1 nears open-source release with a reported 0.007% eval error rate

A developer prototypes isolated VM analysis for suspicious links and files

Daytona sandboxes become available inside Scale AI's AgentEnv

A developer argues that creative agent interfaces should move beyond chat

Google documents how to build and validate Antigravity plugins

Zeeg tests automatic agent feedback through MCP

Gemini Argon leads BallotBench trading but leaves losing positions open

Report: SpaceXSI reportedly adds watermarking to API text outputs

Google adds a security and data-use warning to Antigravity

Gergely Orosz argues SOC 2 alone cannot address AI-agent threats

Paulius Ztins recommends turning agent failures into regression cases

Omar Sar calls for stronger evaluations of open-ended AI work

Grim Prompter adds cited threat-model workflows to coding agents

Prospero coding harness uses deadlines to keep unattended agents moving

FrankenSim uses targeted physics prompts to refine simulation modules

Report: malicious Claude Code extension reportedly persisted after Cursor uninstall

Commentator alleges OpenAI withholds generated Lean proofs

LangChain’s Vatsal Trivedy proposes a live view for long-running agents

Theo compares six models on understanding T3 Code’s architecture

Bubble scores generated apps for visual quality, latency and inference cost

Superdot documents Interactive mode workaround for Claude SDK five-hour limits

Report: OpenAI pricing page lists GPT-6 Astra for Law at $10/$50/$1

A developer builds a Fallout 4 mod with four frontier models and 35B tokens

LangChain guide says retrieval changes cut failed RAG retrievals by 67%

OpenCode users can use server-side compaction for faster context handling

Primal VAE experiment generates native 4096×4096 MOSS Image outputs

Google DeepMind expands PlanAssist to 12 more UK councils

Chamath says about 1,900 Anthropic customers spend over $1 million annually

Claude Duet routes a Linear issue into a project drafting workflow

Reanimate recreates web animations and iterates against visual diffs

Nexus reports a $28,775 AI film with 62 million episode views

Adobe Firefly Boards tutorial structures character design and short animation

T3 Code reports mobile prompts held steady over the weekend

Claude fills a research document directly in its built-in editor

Tester shares unreleased Gemini Carbon action-sequence demo

Herdr reports 30-plus mods and 20 layouts for its agent workspace

Delta v0.22 supports multiple terminals across worktrees

OpenCode maintainer denies weakening its hosted DeepSeek model

Wafer reports ranking third on OpenRouter by daily token volume

Ostris releases Primal VAE in 2×1 and 2×2 variants

Mischief 1 launches on fal with a claimed 5× video speedup

CMU paper tests MaxRL across pretraining and RL

Meta studies weight-shared loops as a way to deepen transformers

Anthropic researchers show poisoned training data can backdoor CLI agents

fx v0.0.16 cuts uncached prompt tokens and shrinks downloads

NVIDIA releases Nemotron 3.5 models for efficient agent workloads

Google runs Gemma 4 TPU inference in orbit

Two AI agents flag 2,035 potential issues across 138 economics papers

Cohere advances a C$4 billion Canadian AI investment plan

OpenAI describes work to reduce inconsistency across inference providers

Nesformer demonstrates NES emulation in roughly 1GB of Qwen3.5 weights

Unified Training applies one objective across model training stages

Two 256GB Mac Studios outperform one 512GB system in local inference

River API runs custom PyTorch loss functions across remote GPUs

Matt Pocock builds a Pixel Watch voice app for sending recordings to agents

Vercel Connect adds webhook triggers for Resend, X, Twilio, Notion, and Telegram

Vercel Labs releases agent-desktop for computer use on macOS, Windows and Linux

Anonymous security team reports DeepSeek 1017 generates 50 minutes of findings

OpenAI introduces Private Intelligence data controls for businesses

V3R1T4S launches a hub with a leaderboard and 1,100-plus research problems

Commentator reports limited access to Anthropic model weights

Sentry cofounder proposes separate charges for processing and retention

Study tests whether research agents can rank experiments by novelty

MiniCPM4-0.5B adds local PII detection to Microsoft Presidio

Agent Bake Off compares model sessions and synthesizes their code changes

OpenAI researchers describe monitoring nearly every frontier-model rollout

Parallel launches hourly web monitoring in Slack

Expand Search launches agent search at $1.50 per 1,000 queries

Mastra standardizes query operations across seven vector databases

Firecrawl adds web search and scraping tools to the Eve agent framework

Practitioner describes asynchronous compaction that reuses cached context

Vectorless RAG uses document trees to retrieve relevant sections

Tester reports a purported Google Omni Flash model with 4K video output

Ryugi guides coding agents from UI artboard feedback to screenshot checks

Google winds down Jules, with reported retirement on December 31

Antigravity adds a plugin marketplace for Notion, Stripe and Figma

John Carmack argues agents need better reusable code building blocks

Cline nearly doubles Nemotron throughput with five vLLM changes

An OpenAI researcher says an 8B run will guide a model 5–10 times larger

Snaptrude Agent turns design briefs into editable building models

Pika releases open weights for H3 Flex with API pricing from $0.01 per second

Kimi launches 3D Make with Blender and CAD connections

Baseten releases joey-1, a 2B decision model tuned for low latency

Posts report GPT Image 2.6 references in OpenAI documentation

OpenBMB open-sources ForgeKernel after an agent beats FlashMLA

Google AI Studio adds XR Blocks for prompt-built 3D experiences

DeepSpeed 0.19.8 fixes a ZeRO-2 gradient accumulation precision bug

NyoomFloat12 proposes lossless 12-bit compression for BF16 model weights

Posts claim new Fable and Opus checkpoints are in testing

SemiAnalysis reports SGLang nears 1,000 tokens per second per user

SkyPilot proposes two rack-local spare nodes per NVL72 rack

Epoch AI estimates chip compute capacity is growing 2.5× per year

AgentBaton reportedly cuts agent API costs by 65.9%

Base Labs finds SFT gains can fade after reinforcement learning

Qwen repository changes reference six Qwen 4 variants and Qwen 4.5 Max

MAST-Data maps failure patterns across multi-agent frameworks

A Grok Bot workflow qualifies inbound leads with reviewed replies

Luna spawns five subagents on a task intended to save costs

AgentWeb exposes website tasks as callable actions instead of browser clicks

Phonic-1 launches in LiveKit Agents with reported tool calling

Cartesia launches Managed Agents at $0.03 per minute

Managed Deep Agents 0.10 adds browser actions in public beta

Switch generates finite action choices for decision models

TokenEconomizer learns repository habits from coding-agent sessions

Mintive gives enterprise agents separate identities, permissions, and budgets

Kernel adds native browser support for Docker agent sandboxes

herdr-core keeps agent sessions running through app restarts

Matt Pocock fixes Codex skills that passed inherited context to subagents

Kimi adds OAuth connections to Gmail, Drive and Calendar

Cua Spaces adds shared desktops with view and edit permissions

Vercel MCP adds structured bug reports and feature requests

Hugging Face adds trace viewing for Claude Code, Codex and Pi

Rootly data shows developers are paged twice as often after hours

Savi describes an agent loop that detects and fixes production bugs

RuHarness shows co-evolving rubrics can improve held-out agent tasks

Grok 4.7 scores lower on slop and sycophancy than earlier releases

SonarQube Hunter Agent finds six access-control flaws in a TypeScript app

Together launches US-only inference endpoints for seven open models

Cognition reports Gemini 4 Carbon scores 40% on FrontierCode 2.0

Epoch reports AI-built Firefox exploits during a public-fix patch gap

OpenClaw v2026.10.9 fixes update rollback and archived agent data

WM-Basin detects distilled models with 79% accuracy at a 0.1% false-positive rate

Modal reports Nova scores 83.9% with Opus 4.8 on Terminal-Bench 2.1

Researchers allege Dodo Payments support-chat flaw allowed email-only impersonation

Jason Weston shares a paper on human-guided research agents

Folio offers a free PDF reader with optional ChatGPT and Claude panels

Deep-ML publishes a hands-on course for building AI agents from scratch

OpenAI updates its Model Spec with more guidance on side effects

Anthropic reports 250 customers spend over $10 million a year on Claude

ElevenLabs releases SFX v3 in ElevenAPI

Baseten and Postman pair API agent tooling with production inference

Serval says its Catalyst workflows replaced ServiceNow in 45 days

Spawn says ten employees shipped 77,000 changes using AI agents

Skyworks uses LangGraph agents to halve root-cause analysis time

Claude mobile apps add multi-account sign-in

OpenAI showcases Codex workflows for science and overnight prototyping

T3 Code adds a “send now” interruption during compaction

GitHub details Copilot agent updates for VS Code workflows

An engineer warns that agent-built infrastructure may trade away guarantees

Capy 0.6.0 adds plugin discovery, themes and faster file sync

GitHub Copilot adds cloud automations and stacked pull requests

T3 Code demonstrates faster navigation with Fast Mode

Claude Code invites developers to propose Mods use cases

OpenCode desktop replaces the terminal for one agent developer

A CI discussion details the hidden work behind build pipelines

Bitrig demonstrates building and testing an iPhone Duo app in its simulator

Atlassian deleted 883 customer sites in 23 minutes with a cleanup script

Developer reports 15× faster CI after moving platform tests to CD

Practitioners challenge Giotto's European sovereignty and benchmark claims

Vikram S. Krishnamurthy shares a conference list for AI infrastructure

GitHub blog ranks highly engaged open-source repositories

Apple's SynthAPI study tests how environment diversity affects tool agents

DeepSeek's FrontierCode results point to broader RL coverage gains

Genspark trains a slide model on MiniMax with Fireworks AI

Databricks says Unity Gateway customers cut token spend by about 30%

A model-tuning study finds human ratings can diverge from benchmark scores

PixelUMM uses raw pixels as a shared interface for vision and generation

Sunra routes video and media models through one key

Matthew Berman says local-model runs shorted his office circuit

ForgeKernel tops FIB tracks in NVIDIA SOL Bench and Huawei CANNBench

Particle launches Particle-1, a token-efficient coding model

SemiAnalysis reports more cached tokens with SGLang DRAM offloading

Farza demonstrates real-time AI-generated graphics on screen

Meta's Muse Spark 1.3 ranks highest for friendliness on AA-PersonalityBench

fal adds VEED Dubbing with 94-language voice cloning and lip sync

Artificial Analysis launches an image-to-video benchmark

Samaya reportedly beats Fable 5.1 at lower cost with Nemotron

ASE demonstrates panel-level packaging for AMD Venice interposers

Tinker and Anthropic’s KV-cache billing models raise incentive questions

A FrontierCode analysis argues RL environments drive coding-model gaps

NVIDIA CUDA resources explain how to inspect compiled GPU kernels

ImagiRep finds spatial reasoning benefits from representations beyond RGB

AMD and Miles report faster agentic RL steps on MI355X

Build-A-Chip-Bench evaluates agents across the chip-design workflow

Alphabet reports Cloud growth and 32 billion model API tokens per minute

Microsoft reports Azure growth alongside data-center spending

Ollama demonstrates local open-model analysis of security scripts

Qwen3-ASR 1.7B reports high throughput for audio transcription

Together says parallel rollouts lower per-run post-training costs

Alibaba’s Z-Image offers a 6B-parameter image model with public weights

The Information reports Chinese labs distilled OpenAI's Astra

ChatGPT shares weekly usage limits across Codex and Workspace Agents

Artificial Analysis v5 widens the measured gap between open and leading models

ResearchBoost-Agent reports a 9.6× training speedup by removing Python overhead

A developer reports shifting coding work from Codex to Opus 5.5

Damodaran values Anthropic at $1.475T and flags compute commitments

OpenUI Lang estimates the token speed needed for responsive generative interfaces

Grok Bot 0.78.1 adds shopping cards and subscription workflows

Perplexity rolls out voice mode for parallel work on desktop

Developers use Git worktrees as a lightweight agent sandbox

A phone text shortcut can invoke skills from an agent repository

Photon brings messaging across iMessage, WhatsApp, and SMS to Capy

ClickUp demonstrates agents using browsers for knowledge-work tasks

AssemblyAI recommends WAV recordings for better voice clones

Jlongster routes code activity and responses through Slack agents

Dzhng argues robotics may need coupled fast and reasoning models

Ben Hylak argues products could expose headless MCP interfaces to agents

Vincent Koc describes using realtime voice to coordinate agent threads

T3 Code adds experimental fleet control through one MCP server

OpenAI previews ChatGPT dots over iMessage or RCS

Suno opens early access to its MCP beta for song creation

Perplexity demonstrates an in-game MCP bridge from GTA V

Remote-computer skill adds isolated Linux VMs on Apple containers

Google previews Connected Apps Platform for Search and Gemini developers

AA-Robotics compares Opus 5.5 and GPT-6 Astra robot motion plans

OpenClaw developers account for 13.8% of Linux Foundation top contributors

ThinkWorld uses imagined state transitions to improve robot reasoning

Rime launches Prism for controllable speech-to-speech agents

Nace releases Drex Max 1.1 with a 262K context window

Pi Durable’s task graph is recommended as a pattern for agent workflows

Theo describes keeping a personal Grok Bot off-screen during videos

Researchers warn multi-agent systems are growing faster than benchmarks

Amp plans shared web profiles for multiplayer computer-use agents

Kylon puts people, agents, and project files in shared workrooms

John Carmack uses a battery-change task to illustrate robot dexterity limits

Timescale shares pg-aiguide for PostgreSQL coding agents

Suno launches an MCP connector for music creation

Vidas Bacevičius says voice teams must test agent dependencies

Browserbase shares bell-stack, a reference architecture for personal agents

Theo says an agent PR was blocked for adding 100 bytes

Tomo filters sensitive personal context from group chats

Coding benchmarks measure distinct agent skills, not one overall ranking

Cognition benchmark chart draws scrutiny over mixed model sources

CI discussion identifies test orchestration gaps beyond build tools

Fable 5.1 edges Kimi K3.1 Pro on the Intelligence Index chart

Researchers report MCP tool descriptions can trick agents into leaking data

ElevenLabs says Cue aligns an hour of audio in about 0.5 seconds

Study finds heart images steer Opus 4.7 and GPT-5.4 toward recursive code

Arena finds fewer repair-loop complaints for frontier coding models

A study estimates AI text appears on nearly a quarter of web pages

FrontierCode discussion argues coding benchmarks measure distinct skills

A researcher compiles a RISC-V CPU directly into transformer weights

Boat CI lets teams allowlist users who can trigger public-repository jobs

A cybersecurity workflow could generate synthetic data for formal-verification specs

MathArena reruns reveal inference-provider errors can distort evaluations

OpenAI shares a Lean-formalized proof challenging the Exponential Time Hypothesis

Log-To-Leak demonstrates prompt-injection exfiltration through MCP logging tools

LiveBench offers a contamination-limited resource for model comparisons

Matt Lam questions why agents still struggle to build evals and harnesses

Swyx says Gemini may lead Artificial Analysis v5 and tests Antigravity

Independent tests find DeepSeek V4.1 Flash mislabels real conversations as evaluations

A post alleges attackers gained administrator access to Anthropic Slack

MiniMax M3.1 Flash tops an automation benchmark at lower cost

Turbopuffer narrows billions of security logs before model analysis

MiniCPM-OCR pairs document parsing with Jev reliability scoring

Theo explains a Linux move with ZFS compression in his workflow

A practical LLM wiki promotes concepts only after repeated sources

TTME uses evolving memory instead of weight updates during math search

Opik guide separates deterministic RAG checks from LLM judging

Harvey rebuilds its platform for FedRAMP Class D certification

Amp adds a keyboard shortcut overlay for quick thread actions

Articul8 introduces Bivia for multimodal workflow decisions

Oracle reports 130,000 ChatGPT users and 95,000 Codex users

Mastra workshop walks through deploying a TypeScript agent with approvals

Cursor engineer describes a year moving work to cloud agents

OpenAI Dot catches a scheduling mistake by monitoring email

Omar Sar argues coding interviews should test agency and judgment

Anthropic engineering manager replaces coding interviews with other signals

Sakana AI opens enterprise early access to its API Platform

Meta gives five universities access to AI models for research

LangChain says schedules drive over 70% of Managed Deep Agent runs

AI models catch planted clinical-trial paper errors faster than reviewers

Gemini Notebook adds hour-long live audio recording on mobile

Gergely Orosz argues autonomy is becoming a baseline AI-era skill

Paper and Cursor turn style references into board-deck drafts

Magnific Move turns image assets into beat-synced motion edits

Vincent Koc shares a phone-first workflow for remote agent work

Paper plans React components that agents can read and edit directly

OpenCode team starts a project for cross-platform native apps

CopilotKit is building a Flutter SDK for agent-native mobile apps

Together AI maps a composable open-source AI stack

Claude Code turns long documents into short Remotion video recaps

Kubernetes control loops inspire an agent workflow for keeping forks current

RepoPrompt CE 1.7 adds granular permissions and session recovery

Yuchen Jiang routes AI-written code through an AI review agent

Entire lets developers claim GitHub usernames for forge migration

Amp releases an Android app for following coding-agent work

A compute-time perfect hash can speed up known-string lookups

Tigris explains how it builds transactional indexes on FoundationDB

T3 Code 0.1.3 cuts thread-loading and connection data transfer

Benchmark reports uzu reaches 92.1 tokens/s, about 4× llama.cpp

SemiAnalysis reports Rubin runs Kimi K3 above 850 tokens/s per user

Stanford experiment pairs an 8× batch increase with 8× weight decay

Microsoft releases MAI-Image-2.7-Preview, ranked second in image editing

AudioFM reportedly leads or ties prior methods on 10 of 12 audio tasks

Lattner reports mobile MAX beats llama.cpp on CPU and GPU

Charm launches 8B Prism to route coding tasks between models

Gemini adds a prompt to check Google-generated media with SynthID

Snowflake shares an overfit-first recipe for scaling agentic reinforcement learning

Neywa Labs targets scalable deployment for on-device AI

OpenRouter operator says one GB300 rack could serve nearly all DeepSeek traffic

Fastino explains reinforcement learning for calibrated decisions

Oracle plans to host Google TPU capacity outside Google Cloud

OM-1 robot carries a box up stairs using glove-recorded demonstrations

SparkLLM releases a 1.7B agent model with one-million-token context

Junie Local runs coding tasks on Windows PCs with RTX GPUs

GenXR-Bench finds coding agents struggle to build interactive XR apps

Cline Desktop adds local model setup through Ollama and LM Studio

Hugging Face adds a CLI view for cached kernel variants

Together AI lets developers post-train MiniMax M3.1 Flash with custom data

Anthropic reports inference revenue per megawatt nearly quadrupled

Yacine's interview explains transformers that predict their next hidden state

Epoch AI estimates the US receives 72% of new AI compute since 2022

H Company reports 24% faster reinforcement-learning steps

Roon argues AI research is harder to automate than mathematical proof

Anthropic secures $63B in chip financing backed by Broadcom

A new accelerator chart compares MI455X, Rubin, Ascend, and Intel specs

Ling 3.1 Flash 1019 improves on its predecessor across evaluations

LearnOpenCV compares Terra and Opus on a CAD floor-plan task

OpenAI plans to buy Google TPUs for its compute build-out

Maya-2 demonstrates natural-sounding speech from an India-based team

Towards Data Science explains why mid-session model routing can cost more

A paper links parameter symmetry to effective learning rates

Cloudflare reports cross-request data exposure in GLM-5.3 serving

Researchers report new experiments on subliminal learning

The Turing Post curates a weekly list of agent and world-model papers

SGLang explains when data-parallel attention can beat tensor parallelism

Wafer’s OpenRouter usage approaches one trillion tokens per day

LangChain’s SFT workflow spends human effort on trace selection

Cartesia Managed Agents scores 91.7% task success in Speech Agent Arena

Skill-diet reportedly cuts agent skill tokens by up to 80%

Nous Portal paid plans add a hosted Hermes Agent computer

CopilotKit releases OpenDuet, a self-hostable proactive agent

Developer runs 24 DeepSeek subagents under Claude Code orchestration

Phonetics helps an agent improve silence detection in audio editing

Explain Interface and Chrome DevTools MCP turn web inspection into a prompt

Thdxr says agents make granular least-privilege keys easier to manage

LangChain open-sources a Slack design agent and explains its harness

ElevenLabs adds a sound-effects plugin for Hermes Agent

MANDATE maps 23 failure types in delegated AI agent work

Box uses a parent agent that spawns child agents for complex requests

Render adds API keys scoped to specific permissions

Entire repositories can now deploy backends through Appwrite

HarnessAgent adds image inputs across its harness adapters

Replit opens Cloud Agents for infrastructure requests and tested PRs

Odysseys2 tests web agents on 100 long-horizon live tasks

Jido frames dynamic interfaces as durable agents that reshape UI state

Jev Router exposes more than 10,000 tools and models through one MCP server

NVIDIA and Meta explore out-of-process monitoring for Muse agents

Mastra lets workflows queue typed signals for the next agent turn

Mastra publishes a workflow template for qualifying sales leads

Black Forest Labs demonstrates FLUX-mimic controlling a real robot

LangChain identifies five building blocks for internal company agents

Cartesia Managed Agents raises speech-task success to 91.7%

Vercel v0 adds editable reports, charts, and HTML artifacts to chat

Arena adds OAuth connectors to Agent Mode for external services

Google and partners ship Kubernetes-native llm-d 1.0

agentloop packages a bounded, traceable agent loop for Go

Air Teams connects repository events to cloud agent tasks

Theo runs 26 T3 Code threads across a fleet of personal machines

Delta CLI lets agents share sessions through a common thread

Amp documents server egress IPs for allowlisting integrations

Infisical documents a sidecar proxy for secrets in Modal agent sandboxes

Dex Horthy warns that subagent hype can undermine context management

Matt Pocock’s skills repository codifies repeatable agent workflows

gnhf runs documented, commit-by-commit autonomous coding loops

Theo reports Cursor-generated changes broke a T3 Code nightly

Will Depue alleges Cathedral gained Anthropic Slack admin access

Axiom says its agents formalized a finite simple groups proof in Lean

Deedy reports OpenAI and Anthropic models found a 2^(n/√log n) 3-SAT bound

Grok Bot 0.77.0 adds permissions for automatic review rules

Developer alleges companies are reselling agent session data

OpenClaw v2026.10.8 fixes stuck chats and missing replies

Hamel Husain and Shreya Shankar publish a book on evaluations

Anthropic study finds experts followed an AI agent's assigned lean 86% of the time

Cursor Research lists five criteria for building useful agent evaluations

Node.js proposal would enable ASLR in Linux x64 binaries

Vercel Observability adds custom threshold and anomaly monitors

Axiom agents formalize the finite simple groups classification in Lean

CodeClimb uses overnight evaluations to optimize code for a dataset

Paulius Ztins argues coding agents need a separate evaluator

Matt Pocock distinguishes fully verifiable agent tasks from partial checks

Infisical keeps projects runnable locally without environment variables

AuthorizationBench tests whether agents stay within user-granted permissions

CodeRabbit commits $1 million to free AI security scans

SidePulse maps JetBrains Junie activity to a physical status light

AxiomProver finds and repairs a gap in a finite-groups proof

A researcher finds a Qwen3.5-4B neuron linked to a speaking style

Artificial Analysis launches Intelligence Index v5 with nine evaluations

AGMAI calls on OpenAI to publish withheld Yang-Mills proof files

SlopCodeBench publishes new findings on coding-agent output quality

MindFort offers hosted DeepSeek-based penetration testing for companies

Omar Sar recommends scaling coding agents while measuring failures

A PR review tool flags changes that could cause production incidents

OpenClaw user reports a loop running continuously for more than a week

Artificial Analysis releases Intelligence Index v5 with ten evaluations

A time(NULL) edge case once disrupted parts of Cloudflare DNS

A Vapache audit finds an RCE path in minutes and maps rising bug complexity

A workflow removes GitHub Actions YAML in favor of native tool configs

A debugging guide recommends replaying failed agent runs before changing models

A harness design separates DONE, RETRY, STOP, and UNKNOWN outcomes

Reflection opens a Beam challenge for demos, models, and datasets

A Qwen3.5 experiment steers an Einstein persona with a sparse neuron direction

Obliteratus organizes researchers to compare model abliteration methods

Anthropic documents how to retry Claude refusals on fallback models

Chandra OCR 2.6 claims faster, cheaper document parsing than frontier models

Sentient OS uses local files to give desktop agents persistent context

Ada reports 77 ms retrieval and 4× smaller vectors with Cohere Embed

StreamVibe studies agents that respond to changing video intent

MICE-P offers compact rerankers with precomputed document representations

Vercel says Agent sessions now cost 67% less on average

A study finds fine-tuned rerankers beat decision models through 9B parameters

Oracle adds image-aware persistent memory for agents

Concurrence builds patient timelines with its Health World Model

Qdrant recommends validating rerankers against tuned fusion first

Jev matches a frontier model on RAG checks at far lower cost

Researchers share an Ettin reranker recipe with Sentence Transformers checkpoints

LlamaParse and Vercel build an interactive reader for Anthropic’s S-1

Grok Bot adds a manual memory-import flow from other assistants

A skill-writing comparison finds imperative descriptions guide agent memory use

Parallel documents a Decisions API with batched questions and citations

Jev powers a macOS dock that switches modes based on user activity

Anthropic extends Claude to state and local governments at 40% off

Steve Yegge describes the Counting House pattern for enterprise AI workflows

Carbon-X approaches frontier models on five DNA understanding benchmarks

Linear adds LaTeX math blocks to its issue editor

AuthorizationBench finds agents exceed user permissions in 19.2% of runs

Meta says 30% of non-engineering product staff now ship code

InteractiveWorld benchmarks multimodal assistants guiding people through tasks

Anthropic reports sharp growth in Claude use across government

Anthropic acquires Mithril Medicine, a one-person drug-development company

Dario Amodei explains why safety disputes led him to leave OpenAI

Rest reports 52% faster sleep onset with its voice AI coach

Nathan Lambert argues that model labs ship their organizational structure

A developer builds a disk cleaner for worktrees, model weights, and builds

Ostris says agent-written fixes can cost less than reviewing outside PRs

Dabit moves most coding work from local machines to cloud agents

Harvey says its API will expose product capabilities to law firms

Superset rebrands as Paradev, an editor for parallel coding agents

Peter Steinberger reports 2–3× gains from Bun's AOT compiler

T3 Code introduces a redesigned interface and cleaned-up settings

Google reports Gemini 4 Argon speeds FarArray graph algorithms up to 180×

Report: Anthropic reportedly delayed pre-training a new frontier model

DeepSeek V4.1 Flash narrows the gap on expert math problems

ARC Prize measures Nemotron 3.5 Super VL at low scores on ARC-AGI

Anthropic directs an estimated 70% of compute spend to internal use

Peter Gostev publishes a chart-based guide to Anthropic's S-1

Cloudflare introduces Clairic for typed extraction from text and images

NVIDIA reportedly lowers Rubin NVLink and HBM4 bandwidth figures

A 320-micron photonic-crystal modulator targets 200G per lane

Unsloth releases MiniMax M3.1 Flash 2-bit quants with 91.5% token agreement

Tencent releases Hy-SimulMT for Chinese–English interpretation with claimed 3-second latency

MCP-Injector finds tool faults cut agent scores by about 60%

DeepMind finds LLMs fail statistical tests of randomness

ElevenLabs publishes a Realtime skill for coding assistants

Yuchenj proposes foldable phones as control planes for agent fleets

Kimi K3.1 Pro scores 76.8 on Browser Use Benchmark

Geoffrey Huntley recommends composing small bots as reactive streams

Goodfire commits $10 million to reverse-engineer a frontier model

Alignment Hub launches a model misalignment incident record

Posts claim Anthropic's internal model disproved ETH using 14 million tokens

Jev screening cuts profile lookup costs to $1.49 for 320 profiles

Google Research demos DEEPSEA across 12 Southeast Asian languages

Matt Pocock recommends splitting risky changes into reversible PRs

A Grok Bot user builds a command center for a fleet of agents

Vercel lets teams choose metrics shown on project cards

Cursor adds DeepSeek V4.1 Flash at $0.60–$0.98 per task

Artificial Analysis launches Code2Video-AA across 114 video briefs

Brian Scanlan tracks customer incidents per 1,000 merged PRs

Thinking Machines finds 80 unique Qwen completions in 1,000 temperature-zero runs

OpenCode developer reports performance issues in indefinite sessions

Fields Model retakes the ARC-AGI-3 lead at 71.48%

Legora cuts legal-agent costs 30% with compaction and custom tool search

Polylane publishes a Git-backed agent memory graph that learns from incidents

OCRia-1 reports leading scores on complex PDF table benchmarks

Nace launches Drex Search to retrieve cited pages from cloud files

Anthropic's S-1 details more than $4 million in founder security costs

Runway launches App Studio for building custom model interfaces

QuixiAI releases open-source Claude Desktop alternative for local or cloud models

Legora runs 2,000 legal-agent evaluations in parallel on Modal