Claude Duet ships as a proactive agent app
Anthropic is productizing long-running agent behavior into a user-facing app that surfaces tasks, keeps working after the user leaves, and pauses for confirmation before taking actions.
Technical AI signals · Daily at 8 PM ET
Wednesday, October 21, 2026
Today split between Anthropic shipping Claude Duet as a proactive agent app, a sharp supply-chain and training-surface security wave, and more concrete agent harness and verification work.
Coverage update: Some YouTube sources were unavailable for this edition.
Anthropic is productizing long-running agent behavior into a user-facing app that surfaces tasks, keeps working after the user leaves, and pauses for confirmation before taking actions.
The exploit claim is unusually concrete: a regularization term can allegedly turn model weights into RAM and SGD into a clock, meaning arbitrary computation could run inside normal training loops.
Its kit turns developer intent into a formal spec and uses Verus to prove generated Rust satisfies it, pushing AI coding from best-effort generation toward mechanically checked correctness.
Two separate results point the same way: fixed models got materially better when the harness or interface evolved from trajectories, with OO-Harness lifting AutomationBench to 77.4% and Agent S4 raising an open 27B model to 21.3% on OSWorld V2.
A compromised LiteLLM release that installed a credential stealer and backdoor is the clearest reminder today that the agent stack now has to be operated like critical infrastructure.
The collaboration covers faster simulation, open benchmarks and datasets for physics AI, and evaluations for agentic AI in CAE and EDA.
It reports 129,514 motions and says the method matches prior state of the art with one GPU in three hours instead of 128 GPUs for seven days.
The method uses one model as both teacher and student, and the post says it matched GRPO accuracy on MATH with 6.5x fewer training tokens.
On a 4 vCPU / 8 GB VM, PGX fits 1,000 untouched databases in 128–155 MB of RAM and uses 0.33–0.43% of one CPU core at idle.
This Hugging Face post describes a 77% inference-cost reduction from specializing SLM vocabulary in real applications.
GitHub says it moved 99.99999% of its repositories to Azure and is using per-reference quorums and object storage to scale the rest.
Humming is a quantized inference kernel library from Ant Group and the vLLM community with broad GPU and quantization support.
vLLM says Flint checkpoints have 40B total parameters, 2.5B active parameters, and about 500 tokens/s on one GPU for a single request.
NVIDIA Sentry is a runtime security system for AI agents that can cross security boundaries while completing tasks.
NVIDIA frames protein design as a search problem over more than 25 billion eight-residue sequences.
The audit reran the LiveCodeBench half of ImpossibleBench on October 2026 frontier models from Anthropic, OpenAI, and Google.
The vLLM team used asynchronous preparation and NIXL EP to enable near-zero-downtime Elastic EP scaling.
Modal added Sticky Sessions so voice agents and multiplayer games can keep strong affinity without losing autoscaling.
OpenAI, PerturbAI, and Scripps Research used distributed AI agents to find leads in data from millions of brain cells.
Together AI says users can post-train open LLMs such as Qwen, GLM, and Kimi with SFT or RL while Together runs the infrastructure.
Modal Dojo is a library for production-grade LLM training with building blocks, tuned defaults, and observability.
vllm-torchtpu adds a vLLM backend for Google Cloud TPUs with Pallas kernels, multi-host parallelism, KV cache offloading, and disaggregated serving.
The post compares decision models with LLM judges for automated evals and says they promise fast, inexpensive classification.
OpenAI added new evaluations for software engineering, SQLite, and COBOL migration.
Commenters say the framework delivered high-throughput, low-latency inference for DeepSeek V4.1 Flash and one cluster reached 155 trillion tokens per day.
Commenters discuss a PHP security release and a three-line nginx mitigation that blocks multipart requests.
Commenters describe DuckDB 2.0 as faster on high-core-count machines and note production use with Iceberg and Arrow Flight.
Commenters focus on transcript leakage between higher- and lower-privileged users in shared agent rooms.
Commenters note that Google Cloud customers can draw down existing spend commitments for OpenAI model purchases.
Commenters describe systems that add algorithmic checks so agent changes stay limited to the intended area.
Commenters argue the problem spans downstream packages and backports, and one asks how fixes can be applied to older versions without upstream security updates.
Commenters describe canary tokens as honeypots for critical credentials and say the same idea can be used across AI infrastructure.
Commenters discuss Chromium-specific rules such as keeping crates tiny, avoiding nested modules, and importing type names unqualified.
Matrix-Game 4.0 separates game execution from visual generation with an executable game harness, a multimodal coding agent, and an autoregressive renderer.
Muse Glimmer is a 30B-parameter model with a perception encoder, supports text and image inputs, and runs locally at about 4-bit precision.
The method reuses discarded action tails and can match or beat a native 10-step Gaussian baseline with fewer denoising steps.
CoEvo-RSI uses a shared evolutionary archive and execution feedback to coordinate harness and model checkpoint updates.
ASAP preserves the completion interface and adapts existing agent scaffolds through structured trajectory continuation for pretraining-time evaluation.
The paper finds models make more errors when tasks are composed and attributes the effect to an implicit thinking budget.
The paper argues that stored memory abstractions can reproduce errors by feeding back into retrieval, reuse, and later execution trajectories.
SpecKCast forecasts skipped diffusion steps by tracking spatial-frequency changes that vary across denoising stages.
The paper introduces FocusGUI, a lightweight adapter for step-adaptive GUI perception that addresses sequential distraction in streaming omni-modal agents.
OverSeer adds fixed auxiliary camera views to help a multi-agent world model keep time consistency and agent consistency across streams.
EdgeKernelBench pairs a unified IR deployment framework with a hierarchical benchmark for edge workloads across heterogeneous backends.
Self-Credit turns sparse RLVR outcome rewards into dense supervision by generating citation-based credit during rollouts and rescaling it within groups.
Sage-A1.5 is a 35B MoE agentic model that adds task delegation, parallel subagents, verification, and open scientific discovery.
Lafa scales a blockwise flow language model to 8B parameters and supports fast inference through prefix caching and blockwise parallel decoding.
Ford Prior says GEICO ran the AI-native war room three times across rentals, damages, and salvage using FleetKit and a markdown rulebook.
The episode includes a segment on a severe PHP vulnerability and on AI models uncovering more security flaws.
Gareth Yoder says exposing the legacy code base through an MCP server cut research time from seven days to 15 minutes.
Tenex keeps plans, project context, and decisions in version control while paying for outcomes instead of hours.
This is Lecture 17 on EKF localization from Stanford’s 2019 course recording.
Airbnb’s Project Everest captures knowledge from employee workflows for agents and evaluates whether the shared knowledge improves engineering work.
The episode walks through an agent loop with approval gates, MCP clients, step budgets, and checks before each tool call.
The guide says to compare models on real workflows by freezing a held-out set, repeating runs, and reporting uncertainty.
Qwen3.8-27B is compatible with Transformers, vLLM, SGLang, and TokenSpeed, and Qwen says a hosted version with 1M context is coming soon.
LTX-2.5 is tagged for image-to-video, text-to-video, video-to-video, image-text-to-video, audio-to-video, and video-to-audio.
EmbeddingGemma 2 is a 740M-parameter multimodal embedding model that maps text, images, video, and audio into a single 768-dimensional vector space.
Timestral-209M supports univariate and multivariate forecasting, up to 32,768 historical steps, and forecasts up to 1,024 steps in one forward pass.
LFM2.5-24B-A2B is a 24B-parameter reasoning model with 2B active parameters and up to 1.7× faster on-device inference than comparable MoEs.
Indra OCR is a 1M-parameter CPU OCR model that reads handwriting, handles 80+ languages, and processes about 44 pages per second on 32 CPU cores.
Hy-Memory-30B-A3B is a memory-writer model with about 3B active parameters that replaces the general-purpose LLM in memory frameworks like Mem0 and MemOS.
FDE-Gym is a benchmark dataset for forward-deployed engineering.
MiMo-V2.6-RL-oss provides RL training environments for coding, cyber vulnerability reproduction, general knowledge work, visual web development, and music tasks.
The model uses a sparse MoE architecture with 501B total parameters, 23B active parameters, and a 1M-token context window.
Drama 3 is a 35B-parameter MoE TTS model with 3B active parameters per token, built for multi-speaker scenes and long-form content.
The Space runs the Drama 3 model inside a @spaces.GPU call for each take and uses Hugging Face ZeroGPU quotas.
The skill runs a six-phase audit, including reconnaissance, coverage-led hunting, candidate validation, structured output, record verification, and target-neutral reporting.
vLLM is an open-source library for LLM inference and serving, with state-of-the-art throughput and efficient key-value cache management.
Orca can run Codex, ClaudeCode, OpenCode, or Pi in separate worktrees and includes mobile notifications plus parallel worktree comparison.
Headroom compresses tool outputs, logs, RAG chunks, files, and conversation history before they reach the LLM, and says 10,144 tokens can compress to 1,260.
ComfyUI is a modular AI engine for content creation.
Kronos is an open-source foundation model for financial candlesticks trained on data from more than 45 global exchanges.
freehire pulls 3.3M+ live job postings directly from company career pages and deduplicates them into one schema.
ai-memory carries long-term coding context across more than twenty harnesses, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and Grok.
Graphify Cloud maps code, docs, PDFs, images, and videos into a queryable knowledge graph and says it can cut token spend by 80%.
Supabase combines hosted Postgres, auth, REST, GraphQL, realtime subscriptions, functions, storage, and AI/vector tooling.
Twenty says teams can define objects, fields, and views as code and ship them to a workspace through the Twenty CLI.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.