Bilingual

Damo Academy unveils an AI agent able to discover superconductors
21 days ago
- Alibaba's Damo Academy unveiled the industry's first AI agent for discovering superconducting materials.
- The AI agent, named Elements Claw, identified four previously unknown superconductors, verified in lab experiments.
- The system screened 2.4 million stable crystal structures in 28 hours and narrowed down 68,000 candidates.
- It was developed in collaboration with Renmin University and the University of Chinese Academy of Sciences.
- Superconductor discovery traditionally relies on trial-and-error due to incomplete theoretical frameworks.
Claude Cowork expands to mobile and web
18 days ago
- Claude Cowork, Anthropic's Claude Code-style agent, is expanding from desktop to web and mobile for Max subscribers, enabling cross-device task management.
- The expansion positions Cowork as an agentic administrative coworker that works in the background, requests human input when needed, and operates across devices, moving beyond a coding tool.
- Anthropic's move reflects a broader trend where AI firms like OpenAI are pushing products into everyday work surfaces, betting on owning the space where work gets done rather than just having the best chatbot.
- Anthropic also launched Claude Tag for Slack, an always-on AI teammate, and emphasizes Cowork's ability to run tasks in the background without an online device.
- Early data from Cowork sessions shows the top use cases are business process operations (33.4%) like reports and checklists, content creation and copywriting (16.4%), with software development accounting for only 8.7%.
- Anthropic aims to make Cowork a reference for integrating AI into daily work, highlighting value in handling 'work around the work' rather than core responsibilities.
Files over tools: how we built our agent with a virtual filesystem and bash
16 days ago
- Knock Agent is an AI agent for managing customer messaging resources, capable of creating and updating workflows, templates, audiences, segmenting users, and answering performance questions.
- The agent was built using a virtual filesystem and bash, inspired by Vercel's approach, allowing it to explore and script on resources via a filesystem instead of exposing numerous API tools.
- A custom Elixir-based virtual bash interpreter called just-bash was developed to provide a lightweight, in-memory environment without the performance overhead of a full sandbox.
- The agent uses a durable workflow with Oban and Postgres, lazy-loading resources into the filesystem and providing tools like bash, read_file, edit_file, write_file, and upsert_resource for operations.
- Skills are introduced to teach the agent resource-specific nuances, and a CLI (knock command) is integrated for querying dynamic data like logs, reducing the need for additional tools.
- Observability is achieved through OpenTelemetry traces to Honeycomb and Braintrust, with a homegrown evals system using LLM-as-judge for testing and monitoring agent performance.
- Future improvements include better session compaction, expanded evals, and exploring subagents for focused tasks, building on the filesystem-based foundation to enhance agent capabilities.
Self-testing AI harness finds its own bugs
9 days ago
- Every morning at 9 AM, Omnideck stress-tests its browser tools on 50 real websites across categories like e-commerce and news.
- When tools fail, Omnideck uses diagnostic tools to identify root causes, such as DOM issues or race conditions, and validates them against its own source code.
- Over a week, this routine found 14 bugs across 35 sites, with examples including content selection logic and modal handling errors.
- Bugs are filed as GitHub issues with proposed fixes, creating a self-improvement loop that enhances tool reliability for future tasks.
- The process leverages existing capabilities like browser tools and code analysis to systematically reduce web chaos and improve performance.
LM Studio Bionic: the AI agent for open models
9 days ago
- LM Studio Bionic is an AI agent for open models designed for tasks like coding, research, and document work.
- It offers local model usage or cloud-based open-source models with privacy controls like Zero Data Retention and no data training.
- Features include voice input with local transcription (e.g., Voxtral by Mistral AI), flexible model execution (local, LM Link, or Secure Cloud), and cost control.
- For coding, it inspects local codebases, explains code, and helps with edits and debugging using models like GLM 5.2 and Kimi K2.7 Code.
- For productivity, it works with documents, PDFs, slides, and sheets in sandboxed environments, offering summarization, editing, and web search.
- It supports both local model downloads via the LM Studio runtime and cloud inference with transient data processing.
- Bionic is a separate app from LM Studio, requiring an LM Studio account for cloud models, and focuses on user control and privacy.
Show HN: Agent in 9 Lines Python
3 days ago
- The script is a minimal AI agent implemented in 9 lines of Python.
- It uses a custom tool to execute shell commands via subprocess based on model requests.
- The agent interacts with the user through command-line input, sends conversation history to a GPT-5.6 model API, and displays responses along with token usage.
How we verify Cleric’s production fixes
3 days ago
- Production verification is necessary but difficult due to delayed feedback and unreliable self-assessment by the investigation agent.
- A separate 'auditor' agent critiques the investigation process, improving acknowledgment of uncertainty.
- Verification uses a separate agent per issue that reads findings but forms its own judgment from production evidence.
- Dynamic code generation creates checks that run periodically, with the agent re-invoked only when status changes or monitoring ends.
- The verifier focuses on deeper signals (defects) rather than just alert recovery, and uses a rubric refined through failure cases.
- Learning from experience includes avoiding compounding errors from the verifier's own memory and tightening the standard of proof.
- Outcome data shows most fixes are applied on threads without human reply, and recurring alert clusters often stop after a fix is applied.
- Counterfactual analysis and user actions provide labeled data to improve suggestions and confidence scoring.
- The approach generalizes beyond fixes to verify any code or configuration change in production.
How Gumclaw Works
3 days ago
- Gumclaw is an autonomous AI agent that runs Gumroad operations including support, engineering, finance, risk, and a public X account.
- It uses the Hermes agent framework with the Fable 5 model on a dedicated MacBook Pro (M3 Max, 128GB RAM, 1TB SSD).
- Tools include production console, GitHub, Helper (support), X API, web+terminal, and memory files.
- It has ~40 Hermes skill packs and ~100 operational scripts for various tasks.
- Its job functions include support triage, engineering loop, X mentions watcher, finance bookkeeping, and risk escalation.
- It learns without retraining via daily logs, ledgers, policy rules, and live verification, and operates under strict rules.
I gave Pi one tool
11 hours ago
- The idea started with giving Pi a single tool called fabric_exec that accepts TypeScript code, allowing the model to write small programs against a typed API.
- The model's programs return only the necessary data, reducing context usage and enabling real control flow like loops and parallel execution.
- The system evolved to handle schema errors, state management, deterministic compaction, memory, and cross-session recall, with the transcript as the source of truth.
- Fabric was used to build itself: 337 commits in 12 days, 49,569 source lines and 32,059 test lines, most written by the model through the tool.
- Real-world tests included browser automation with Chrome DevTools Protocol and a 3D printing project that lasted eight days, successfully handling context compaction and failure recovery.
- The runtime cost about $1,500–$2,000 in equivalent API usage for development, with 11,155 programs executed and 93.4 million input tokens consumed.
- The key insight: giving the model a programming language inside the agent harness works better than a catalog of tiny verb tools, with strict security via sandboxed execution.

Hasty Briefsbeta

ai agent

Damo Academy unveils an AI agent able to discover superconductors

Claude Cowork expands to mobile and web

Files over tools: how we built our agent with a virtual filesystem and bash

Self-testing AI harness finds its own bugs

LM Studio Bionic: the AI agent for open models

Show HN: Agent in 9 Lines Python

How we verify Cleric’s production fixes

How Gumclaw Works

I gave Pi one tool