#coding-agents

Databricks developed SWE-Lancer, a benchmark to evaluate coding agents on real-world tasks from their multi-million line codebase. The benchmark includes unit tests and manager-grade task grading. Results show that current frontier models complete under 28% of tasks, highlighting significant room for improvement in AI-assisted software engineering.

Fizzbee.ai is a new open-source app that uses formal verification for requirements engineering. It asks follow-up questions, converts prompts into formal specifications to uncover gaps, generates validation scenarios, and produces a specification document for coding agents, aiming to produce working code in fewer iterations.

The article reverse-engineers token pricing of AI coding agent plans (Cursor, Windsurf, Copilot, Codeium). Pro plans are often 1.5–3× cheaper per token than buying the model directly, while Ultra plans can be 10× more expensive. It offers a framework for developers to compare plans objectively.

The article critiques the overconfidence of AI coding agents in assessing the safety of their own code changes, arguing that current LLM-based tools tend to incorrectly reassure users that modifications are safe even when security risks exist.

Simon Willison reflects on a talk by Geoffrey Litt at AIE, who argued that when collaborating with coding agents, developers must understand the code deeply enough to remain active participants in the creative process, avoiding cognitive debt from drifting understanding.

ctx is a Rust CLI that ingests coding agent transcripts into a local SQLite database for search. It helps agents reference past sessions to avoid repeated mistakes and can generate clean transcripts for sharing.

LoopFlow is a tool that lets users write loop logic in plain English, which then runs coding agents to iteratively fix and retry code until all tests pass.

Matrix OS is an open-source operating system for cloud VPS computers, adding a visual UI accessible from web, mobile, desktop, and CLI. It was started at an Anthropic hackathon and reached the top 20 finalists. The project offers both a self-hosted version and a managed service.

Skill Federation is a free search engine designed for AI coding agents, not humans. It fetches relevant skills for tasks using keyword enrichment and reranking, and uses Cisco and Nvidia security scanners. The project was created to demonstrate that AI solutions need a finite set of interventions to perform well in specific applications, achieving a ~30% relative performance lift in tests.

Reap is a framework for automatically curating coding agent benchmarks from real-world GitHub issues, reducing the need for costly manual creation. It filters candidate issues, generates reproducible environments, and evaluates agents using natural-language reference solutions as a proxy for ground-truth tests.

bb is an open-source agentic development environment (ADE) that works across multiple AI coding providers like Claude Code, Codex, and Cursor. Unlike many terminal-based tools, it offers a polished interface where coding agents can spawn new threads, read each other's output, and modify the UI. It is local-first, scriptable via external agents or cron jobs, and designed to be used by itself.

The website presents a survey asking developers which coding agents (AI-powered tools that assist with programming tasks) they currently use, aiming to gather data on tool popularity and usage patterns in the software development community.

A developer created Secret Shuttle, a tool that allows coding agents to use secrets (like API keys) without the agents themselves being able to see those secrets. This helps maintain security when giving AI coding assistants access to sensitive credentials.

The article explores the rising costs of AI coding agents, which can reach $10,000 per week per developer. It critiques the trend of using expensive, autonomous coding tools and suggests more cost-effective alternatives, including using simpler AI assistance or traditional development methods to achieve similar productivity gains without excessive spending.

Kage is a framework that adopts Google's Open Knowledge Format (OKF) for agent memory management. It handles the timing, structure, and verification of memory creation and recall, going beyond Google's standard which does not cover these aspects. Designed for team coordination on shared repositories, Kage works with Claude Code and other coding agents to automate memory maintenance.

Petabyte-scale durable storage has been added to Celesto sandboxes, supporting coding agents, harnesses, and large file storage for AI agent environments.

World Model MCP v0.10.0 is a cross-runtime memory system designed to work across seven different coding agents, enabling shared context and persistent memory between them.

The article critiques the high cost and limited effectiveness of AI coding agents, arguing that many teams overspend on them due to poor software architecture, lack of test coverage, and over-reliance on AI. It suggests that investing in better engineering practices, documentation, and system design can reduce the need for expensive AI tools and produce better results.

Kage is a framework that adopts Google's Open Knowledge Format (OKF) for agent memory, adding verification, timing, and freshness management that OKF alone lacks. It works with coding agents to automatically decide what to save, when to save, and how to recall relevant memory, with best support for Claude Code.

The blog post describes the experience of using AI coding agents to add RonSQL support to RonDB. It details the process, challenges faced, and how agents assisted in implementing the SQL layer, concluding with reflections on the effectiveness and limitations of current AI tools for complex database development tasks.

Assembled open-sourced 143.dev, its internal system for running coding agents as shared team infrastructure. The MIT-licensed platform runs Codex, Claude Code, and others in sandboxes, integrating with GitHub, Slack, and more for centralized control and introspection.

shot-scraper 1.10 introduces a "video" command that uses Playwright to record demos of web app routines defined in a YAML storyboard file. The feature enables coding agents to automatically produce video demonstrations of their work.

Paneflow is a cross-platform desktop application built with GPUI that enables parallel coding agents to work simultaneously, providing a multi-panel interface for developers to manage multiple AI-assisted coding workflows in one window.

Andrew Ng describes three loops for AI-assisted product development: the agentic coding loop (AI iterates on code), the developer feedback loop (human steers product), and the external feedback loop (user testing informs vision). He notes engineers are taking on product management roles as AI accelerates development.

A user on Hacker News asks for help finding a well-maintained, secure sandbox wrapper for running coding agents, recalling a recent project post. They note that while they could build their own using systemd-nspawn, Kata, or bspawn, searching GitHub has become difficult due to AI-generated projects.

Agentic Orchestrator is a terminal tool that runs coding agents through a deterministic engineering workflow (requirements, design, planning, implementation, review). It allows human oversight in early phases and autonomous execution later, with merge conflict and review comment utilities. The open-source tool supports OpenCode, Codex, or Claude Code on macOS, Linux, and WSL.

The website "Tau" teaches how coding agents are built, offering resources and tutorials on constructing AI-driven software development tools.

Researchers from the University of Washington introduce TraceLab, a framework for characterizing coding agent workloads to improve LLM serving. By analyzing real-world agent interactions, TraceLab identifies key workload patterns that can inform better resource allocation and performance optimization for serving large language models.

PMB is a local memory tool designed for coding agents that visually indicates when the stored memory is being accessed or used.

Fame is an external memory and tool-safety gateway designed for local coding agents. It enables agents to securely access tools and retain context beyond their immediate runtime, acting as a middleware layer for enhanced agent capabilities.