ai agent

Bilingual

Show HN: How to build and self-host a code review agent
2 months ago
- Tilde is a cloud-based harness SDK for building AI agents quickly, designed to overcome sub-optimal outcomes from single-harness solutions and improve developer experience.
- Key components include Tools (MCP server integration, credential management), ChatKit (Slack, GitHub integration), and Memory (skills registries, wikis for autonomous updates).
- A practical example is a code review agent that can be tagged in GitHub PRs, securely checkout code, analyze diffs based on system prompts, and comment back with inline and summary messages.
- The architecture allows developers to focus on the core logic (e.g., code review) while Tilde handles infrastructure, reducing setup time to under 5 minutes.
- Deployment is portable via Vercel and Tilde, with a tilde-state.yaml file for easy configuration of integrations like GitHub and Modal sandbox.
I gave Pi one tool
2 months 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.
How Gumclaw Works
2 months 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.
How we verify Cleric’s production fixes
2 months 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.
Show HN: Agent in 9 Lines Python
2 months 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.
LM Studio Bionic: the AI agent for open models
2 months 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.
Self-testing AI harness finds its own bugs
2 months 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.
Files over tools: how we built our agent with a virtual filesystem and bash
2 months 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.
Claude Cowork expands to mobile and web
3 months 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.
Damo Academy unveils an AI agent able to discover superconductors
3 months 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.

ai agent

Show HN: How to build and self-host a code review agent

I gave Pi one tool

How Gumclaw Works

How we verify Cleric’s production fixes

Show HN: Agent in 9 Lines Python

LM Studio Bionic: the AI agent for open models

Self-testing AI harness finds its own bugs

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

Claude Cowork expands to mobile and web

Damo Academy unveils an AI agent able to discover superconductors