Lately, I've been seeing more and more projects on GitHub that help launch multiple versions of Claude Code and coordinate their work.

Crystal - Multi-Session Claude Code Manager
https://github.com/stravu/crystal
Crystal is an independent project by stravu – a desktop application (built with the Electron framework) that allows you to work with many Claude Code instances simultaneously. Each agent session runs in an isolated Git worktree. This ensures that changes made by the AI do not affect the main code until the developer decides to integrate them.

https://www.youtube.com/watch?v=vGwxhBR81zY

Crystal works with any technology stack (Python, TypeScript, Electron, etc.) and integrates with existing projects. It's built around Claude Code because the developer considers it the best and most cost-effective agent on the market, especially for intensive token usage. Using Crystal significantly speeds up the development process – Crystal itself was created in 2-3 weeks.

Claude Code Subagents
https://docs.anthropic.com/en/docs/claude-code/sub-agents
In response to requests, Anthropic itself has added the functionality to run multiple sub-agents. You can create agents for code security review, test generation, database migration, etc. – configured using simple Markdown files that can be stored globally or at the project level. Anthropic recommends generating initial versions of sub-agents using Claude itself.

Each sub-agent works with its own isolated context and set of tools. Instead of one universal AI assistant, developers get a team of specialized agents, each with their own expertise.

#claudecode

Comments

Qwen3-Coder
https://qwenlm.github.io/blog/qwen3-coder/
The Chinese Qwen team, behind the development of advanced AI models, announced the release of Qwen3-Coder. The Qwen3-Coder-480B-A35B-Instruct model uses a Mixture-of-Experts architecture with 480 billion parameters (of which 35 billion are active), supports a context window of up to 256k tokens out-of-the-box, and can be extended to 1 million tokens. Other sizes are expected to be released.

During the post-training phase, the Qwen team scaled up reinforcement learning for code (Code RL), focusing on real-world tasks where execution success is easily verifiable. Additionally, they introduced Long-Horizon Reinforcement Learning (Long-Horizon RL or Agent RL) to teach the model to solve complex engineering problems, such as SWE-Bench, through multi-step interaction with the environment, including planning, tool use, and feedback acquisition.

The model can integrate with Claude Code and Cline.

https://qwenlm.github.io/blog/qwen3-coder/
For interaction with Qwen3-Coder, the developers introduced a command-line tool CLI – Qwen Code, which is essentially a Chinese copy of Gemini Code.

We get performance at the level of Claude 4 Sonnet, only significantly cheaper.
https://openrouter.ai/qwen/qwen3-coder

#qwen

Comments

Zed without AI
https://zed.dev/blog/disable-ai-features
Finally – it's getting annoying that AI is getting into all coding tools :) Now Zed allows you to completely disable these unnecessary features via the settings.json file.

Many companies prohibit the use of AI tools. Some professionals have ethical, philosophical, or environmental reservations about using AI and prefer full control over their code and workflow without unwanted suggestions or AI interference.

GitHub Spark
https://github.blog/changelog/2025-07-23-github-spark-in-public-preview-for-copilot-pro-subscribers/
MS has released GitHub Spark, which allows you to turn an idea into a ready, deployed application in a matter of minutes. It creates a full repository on GitHub with GitHub Actions and Dependabot, ensuring full synchronization.

Spark will independently build a complete application, including both frontend and backend. All this works on AI models such as Claude Sonnet 4. Hosting, deployment, and integration with GitHub authorization – everything is included "out of the box". It is possible to integrate requests for models from OpenAI, Meta, DeepSeek, and others, without the need to manage API keys.

Currently available in public preview for Copilot Pro+ subscribers.

Terragon
https://www.terragonlabs.com/
Claude Code in the cloud, which works as parallel background agents in a separate sandbox, but with access to GitHub repositories. Start or manage tasks from the web dashboard, Terry CLI, GitHub comments, or your phone — work from anywhere.

#zed #githubcopilot

Comments

AWS Kiro v0.1.0 (preview)
https://kiro.dev/
A new "agent-based IDE" (like Cursor and Windsurf) from Amazon Web Services (AWS), presented on July 14, 2025. Built on Visual Studio Code. Pleasant visual theme. Supports MCP. Currently uses only two Claude Sonnet models 4.0 and 3.7.

Available for free in preview, and includes limits that allow you to try the product without interruptions. In the future, various pricing plans are planned: Free, Pro, and Pro+.

https://www.youtube.com/watch?v=Z9fUPyowRLI

The key feature is a mechanism that helps transition from rapid prototyping ("vibe coding") to creating full-fledged, production-ready code ("viable code"). Kiro creates detailed structured artifacts: requirements documents (requirements.md), design documents (design.md), and task lists (tasks.md).

There are also Hooks, which configure event-driven automation, such as updating documentation or generating tests when a file is saved. And Steering, which allows defining description files to guide agent behavior, adding context, standards, and desired workflows.

Kiro's developer, NathanKP, actively interacts with users in the discussion https://news.ycombinator.com/item?id=44560662. Many people express disappointment that Kiro is another VS Code fork – reporting high CPU and RAM consumption.

Some believe that Kiro is "late to the party," as the market has already shifted to terminal agents like Claude Code. I also thought it would be a console agent.

Currently, for me (compared to Cursor and Windsurf), it runs quite slowly. Also, I haven't found checkpoints – you can only undo individual agent edits or last batch.

#kiro

Comments

Grok 4
https://openrouter.ai/x-ai/grok-4
Elon Musk's xAI company has released the next version of its large language model. In their presentation, they used one of the ways to manipulate statistics, by cropping one axis to make the difference seem much more significant. Possibly other descriptions of the model are not straightforward either. But this does not negate the fact that both the model and its agentic mode with 4 competing agents show outstanding results (and pricetag:).

The context window size of 256k tokens is average by current standards. It handles confusing questions and planning very well. Judging by reviews, in practical programming tests it shows average results. It appeared in Cursor's list of available models, but not yet in Windsurf.

Kimi K2
https://moonshotai.github.io/Kimi-K2/
Unexpectedly, a bigger piece of news for programming was the appearance of a new open model from Moonshot AI. The model is not "thinking", but simply a Mixture-of-Experts (MoE). It can work as an agent because it knows how to use tools and functions.

The model can be connected via https://github.com/sst/opencode or any VSC plugins compatible with openrouter. It is also deployed on Groq.

Context window 131K. In tasks where planning is not needed, but simply code generation, the model outperforms Claude Sonnet 4. On openrouter there are different providers.

#newllmmodel

Comments

It is unclear if current AI programming tools can sustain themselves with the current low subscription prices.

Windsurf
The planned acquisition of AI startup Windsurf by OpenAI for $3 billion did not happen. Instead, Windsurf entered into a licensing agreement with Google (this is not an acquisition of the company, but access to technologies), with key Windsurf employees moving to the Google DeepMind team.

One of the main reasons for the deal's collapse was disagreements related to the relationship between OpenAI and its largest investor, Microsoft. Windsurf expressed concerns about Microsoft's (owner of Github Copilot) access to its intellectual property under the existing agreement between Microsoft and OpenAI.

The move of key Windsurf employees to Google DeepMind is seen as a significant reinforcement for Google in the field of AI development, especially in light of their work on the Gemini line.

https://windsurf.com/blog/windsurfs-next-chapter
Cognition, known for creating Devin – the "world's first" autonomous software engineer agent, acquired Windsurf. The Windsurf team will join Cognition. Windsurf's unique intellectual property (IP) will be integrated into Cognition products.

The combined product will allow engineering teams to plan tasks in Windsurf with a deep understanding of the codebase from Devin, delegating parts of the work to Devin agents.

Cursor
https://cursor.com/blog/june-2025-pricing
Starting June 16, 2025, Anysphere (owner of Cursor) quietly changed the terms of the basic Pro plan ($20/month), adding usage-based billing. Many users quickly exceeded the limit (especially when using new Claude models) and faced unexpected additional charges as they didn't fully understand the new terms. This caused a wave of complaints on social media.

https://www.youtube.com/watch?v=Q4aaynNs59Q

Anysphere explained the changes by stating that the new models have become significantly more expensive, especially when performing complex tasks that require a large number of tokens. The company plans to refund users who were charged unexpected amounts and promises to be more transparent in the future.

Replit
Users of another popular tool, Replit, have also faced increased costs for running large tasks.

#windsurf #cursor

Comments

git-ai
https://gitai.run/
The project helps automatically save prompts and track AI-generated code. It is currently in the Preview development-testing stage. It's Git-native, storing authorship logs in JSON format within Git, tied to commits. Written in Rust (fast, cross-platform).

Uses "checkpoints" (git-ai checkpoint) before and after file changes by agents to mark authorship (human or AI). Commands include git-ai checkpoint for marking and git-ai blame <file> for viewing.

Can be integrated into Cursor and Claude Code.

Comments

Creating the Tower of Time game
https://github.com/maciej-trebacz/tower-of-time-game/blob/main/PROMPTS.md
In cases where the code for a repository is created by AI, it would be good to add the entire chat protocol or at least the prompts that were used.

The developer (maciej-trebacz) acts as the architect and team lead: he sets tasks, checks results, points out errors, and directs the process. The AI does not invent the game "from scratch" but implements specific, clearly defined tasks.

Development is conducted iteratively. No complex feature was created with a single prompt. When the developer asks to create a WaveSystem "by analogy with EnergySystem.ts", the AI performs excellently because there is a ready pattern. The prompt "Refactor BasicTower to extract common logic for all towers" is an ideal task for AI.

Conclusions:

• Write as specifically and step-by-step as possible. Bad: "Make enemies." Good (as in the file): "Implement enemy movement. It has a target position. On each update(), it should move towards it. During movement, play the animation corresponding to the movement vector." For the complex RewindableSprite mechanic, the developer described an entire algorithm with several points.
• Engage AI in creative tasks. Ask not only to write code but also to critique the scenario or suggest ideas.
• It is important to provide context. Always specify the files you need to work with using the @file.ts syntax. This is critically important for accuracy. Specify how the system should NOT work (e.g., "towers do not shoot during time rewind").
• Use existing code as an example. Ask it to make new features "by analogy with..." to maintain a consistent code style. Don't just say "there's an error here," but send the error text and logs.

#prompts

Comments

Claudia
https://claudiacode.com/
A Graphical User Interface (GUI) for managing Claude Code - replaces terminal work with a convenient visual interface. The project is free, open-source for Windows, macOS, Linux. https://github.com/getAsterisk/claudia

Provides a overview of projects/sessions/agents, creation of custom AI agents, visual monitoring of costs (tokens/funds), built-in prompt editor (Markdown), chat version control (like Git).

SuperClaude v2
https://github.com/NomenAK/SuperClaude

The configuration framework has been updated, extending Claude Code with specialized commands (18 total), cognitive personas, and development methodologies. The 9 personas can now be launched via command-line flags (--persona-architect, --persona-security, etc.)

MCP integration with Context7, Sequential, Magic, Puppeteer is supported.

Trae Agent CLI
https://github.com/bytedance/trae-agent
In addition to their IDE Trae, Bytedance has also released an open-source agent that works in the terminal. It is currently in an alpha-experimental development stage and is more for testing and improvement.

They emphasize that their agent has a modular architecture. It constantly writes detailed logs and provides a short summary at each step. It supports OpenAI, Anthropic, Doubao, Azure, and OpenRouter APIs. It uses trae_config.json for configuration.

Amazon is also preparing something in the field of a CLI agent under the code name Kiro - we'll wait and see.

#claudecode #trae #autonomousagents

Comments

Build apps with Gemini
https://aistudio.google.com/apps
Inside Google's AI Studio there's a section where you can create simple AI applications using prompts. It's a quick tool for prototyping and testing ideas.

The environment uses the Gemini SDK, runs the app in the browser within a sandboxed iframe, and does not support Next.js, Svelte, Vue, or Astro. It also opens an editor with access to the code. There are examples, including music generation.

You can further deploy it on Cloud Run.

https://www.youtube.com/watch?v=YEzjVvE44R0

The first version of the app generates images with unreadable, distorted text. The author doesn't give up and directly in the AI Studio editor provides the artificial intelligence with additional instructions so that it first analyzes the provided information and then creates a higher-quality result. After this, the app starts generating significantly better creatives with a relevant image and correct text.

#google #gemini

Comments