Google open-sources Harness² for self-improving agents
Harness² is the clearest new agent artifact today: it lets agents rewrite prompts, tools, memory, and skills, then test each change on the actual task without retraining or ground-truth labels.
Technical AI signals · Daily at 8 PM ET
Sunday, October 18, 2026
Today was an operator-heavy day led by Google’s open-sourced Harness² self-improvement loop, a fresh wave of agent memory and plugin plumbing, and more concrete work on evals, verification, and serving performance.
Harness² is the clearest new agent artifact today: it lets agents rewrite prompts, tools, memory, and skills, then test each change on the actual task without retraining or ground-truth labels.
Cursor is turning cloud agents into a more shareable production surface, with interactive canvases plus beta environments that support up to 300 repos and load large setups 5x faster.
DeepSeek says the new Flash release improves reasoning, coding, agentic tasks, and visual benchmarks while cutting task time and output length, which is exactly the kind of serving-side model update operators can use immediately.
The practical point is not just speed: vLLM describes agent-assisted kernel work alongside correctness tests and reproducible benchmarks, showing how model-serving teams are using agents inside the optimization loop without dropping rigor.
Taken together, these talks push the same operational pattern: trace runs, mask PII, build eval datasets, and make agents prove completion with health checks, snapshots, judges, and CI gates before shipping changes.
It lists instant crawling, seven checks across market and holder data, contract x-ray parsing, and a launch radar that auto-scans new tokens at birth.
The post says $30,000 in Claude credits were used to research the author's own tumor.
Lighthouse lets agents share one memory across Claude and ChatGPT, and the Virtuals app also adds one-tap portfolio buying from an X post or screenshot.
Vangrid turns real-world captures into spatial data for robots and physical AI, using distributed contributors instead of centralized data collection.
The open-sourced system uses Claude Code to build a profile, scrape job boards, rank openings, and generate applications after the developer landed a role as an AI engineer.
ADEMark tests 16 ADEs on the same tasks and machine, measuring tokens, speed, and resource usage across frontier models.
The post studies how LLMs aggregate subjective evidence such as 'great,' 'good,' and 'terrible' into answers.
The author fine-tuned a toy model on TinyStories to emit marker tokens tied to subsets of its training data for monitoring.
vLLM added support for Reflection's Beam-1.0, with 501B total parameters and 23B active.
Modal's VM Sandbox filesystem work speeds access by speaking in blocks instead of files.
The paper introduces FBS Debate, which gives worst-case correctness guarantees and improves on the prior debate protocol.
Hugging Face says the pretraining effort uses 2.8 trillion tokens and an open data generation pipeline.
Google Research says the work lays out building blocks for autonomous wet labs.
The commenters say Claude now reviews most changes continuously before a human looks, instead of waiting for a PR gate.
NVIDIA says RTX Spark lets developers code, test, and deploy on one machine while agents write most of the code.
Commenters discuss a code release for HybridAgent and ask how the approach differs from other transformer architectures.
Commenters focus on hosted search APIs, with one suggesting personal search tools like searxng-ai-kit as a cheaper alternative.
Commenters describe Sphinx as a way to check AI-generated code before pushing to prod or requesting more human review.
Commenters say verified kernels make Lean easier to use with AI and speculate about formal verification workflows.
Commenters describe a DIY sub-micrometer XYZ manipulator and mention potential use for diagnostics carts and piezoelectric stick-slip actuators.
Commenters say the simulator could help diagnose and fix C64 boards and test timing edge cases on specific board revisions.
Commenters debate whether the article's requirements make the claim realistic and say the explanation is easier to follow with Claude.
Commenters say sealed classes help with access control and exhaustive pattern matching in Java, especially with dependency injection.
Commenters recommend Python's built-in re.Scanner for cases between a single regex and a full parser, and note parser combinators require more mapping and unwrapping.
Commenters point to Luarocks as a large Teal codebase, and one commenter says Algernon ships with built-in Teal support.
Commenters say the 8087 could handle 80-bit floating point and that early Win32 versions did not save 8087 registers on context switch.
MJev-Spatial turns spatial tasks into a unified classification framework for relations, distances, and 2D point placement.
DeltaStreamer records visual changes in structured memory, then uses query-driven planning and replay to recover evidence for later questions.
Intern-Decision uses a JSON skeleton with one decision placeholder per field and returns structured choices plus candidate probabilities in a single forward pass.
The paper proposes market-based post-training, where agents bid for participation and receive local training credit from collective outcomes.
MaLoW consolidates streaming video history into fixed-size LoRA weights through a recurrent HyperNetwork instead of keeping memory in the input context.
Data World synthesizes environments, tasks, and rubrics from explicit mechanisms, then audits whether the target can be recovered from the observed files.
EDEN adds anchored input perturbations to broaden rollout variation for reinforcement learning with verifiable rewards.
RPO turns attention routing into a stochastic exploration action inside the generator instead of perturbing the reverse denoising trajectory.
OmniCompass splits evaluation into OmniCompass-VL and OmniCompass-AV, with tasks for temporal grounding, spatial grounding, and event understanding.
InfiniSplat v2 freezes the geometry branch, aligns Gaussians with ground-truth camera poses during training, and adds large-angle target view supervision.
The paper adapts frequency-band boundaries and band widths to each mixture under a fixed band-token budget.
ViRA keeps the planner architecture and inference cost unchanged while aligning visual foundation model representations for end-to-end driving.
OneWM uses a shared video-action architecture to treat actions as both conditions for world prediction and targets for policy learning.
The episode discusses Anthropic interpretability work on models that plan many tokens ahead and form general representations of concepts.
The lecture covers information extraction and classic visual recognition in Stanford's 2019 robotics autonomy course.
Jan Curn says agent payment standards such as x402 and AP2 exist, but discovery, wallets, and API keys still block real usage.
The tutorial claims Gemini 4 Argon has a 1,048,576-token context and shows how to access it through Sim.ai and Vercel AI Gateway credits.
The video asks what happens when an agent times out on an API call and nobody knows whether the operation succeeded.
The demo uses FalkorDB to investigate microservice incidents by following connected data instead of relying on 0-hop text retrieval.
MiniCPM-V 4.7 is a 35B-A3B model and the first open-source MoE model in the MiniCPM-V series with smooth on-device inference.
EmbeddingGemma 2 maps text, images, video, and audio into a single 768-dimensional vector space for on-device search, RAG, classification, and clustering.
The model supports image-to-video, text-to-video, video-to-video, image-text-to-video, audio-to-video, text-to-audio, and video-to-audio.
LFM2.5-24B-A2B has 24B parameters with 2B active at a time and is built for on-device reasoning and tool use.
Baidu describes Unlimited OCR as a one-shot, long-horizon parsing model that now supports training with ms-swift and inference with vLLM.
No descriptive text was provided beyond the model tags.
Qwen says the model weights work with Transformers, vLLM, SGLang, and TokenSpeed, and a hosted version with 1M context and built-in tools is coming soon.
The dataset is tagged as a benchmark for agents and forward-deployed engineering.
The corpus snapshot covers 3,148,796 arXiv papers and includes metadata, version history, submission files, and rendered documents.
MiniCPM-V 4.7 is a 1B variant that extends single-image, multi-image, and video understanding with on-device inference.
The benchmark uses one unstructured state input and asks five typed questions, with every answer expressed as a probability distribution.
The dataset provides RL training environments for LLM agents across code, cyber, general knowledge, visual web development, and music tasks.
OpenCode is an open-source AI coding agent with desktop apps for macOS, Windows, and Linux.
Anthropic's plugin directory separates internal plugins from third-party plugins and installs through Claude Code's plugin system.
Cursor's plugin repo includes tools such as teaching, continual-learning, and cursor-team-kit, each packaged as a standalone plugin directory.
ai-memory carries long-term notes across more than twenty harnesses, including Claude Code, Codex, Cursor, Gemini CLI, and OpenCode.
Orca runs Codex, ClaudeCode, OpenCode, or Pi side by side in isolated git worktrees and includes a mobile companion for steering agents remotely.
CLI-Anything provides a hub for installing community-built CLIs and uses generated harnesses for agents to produce artifacts such as CAD builds, 3D scenes, and diagrams.
VoiceStudio supports voice cloning, video dubbing, dictation, transcription, and audiobook creation in 646 languages.
Android Skills packages modular SKILL.md instructions to help agents handle Android workflows where evaluations show weaker performance.
Jasna restores mosaics in videos and images, leaving the rest of the picture untouched, and runs on a GPU.
ComfyUI is a modular AI engine for content creation.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.