AI Agent That Knows When Not to Guess
A hackathon post on building an agent with Qwen and MCP that avoids guessing. The example is a payment of exactly half an invoice from a known customer email.
The weekly · Week 29 · 13 Jul – 19 Jul 2026
AI coding agents: what changed
This week covers MCP and agent tooling, the token cost of coding agents, and sandboxes for running them.
8 picks
One MCP Server Called From Three Agents
A beginner tour of the Model Context Protocol: one Python server that generates images with Gemini, called by Claude Code, a Google ADK agent and a Rust CLI.
Migrating a Production Agent to GPT-5.6
A team reports moving its production agent to GPT-5.6, which ran 2.2x faster and 27% cheaper. After four months, it was the first model to beat Claude Opus in their evals.
Claude Code Sends 33k Tokens, OpenCode 7k
Token overhead of Claude Code and OpenCode measured at the API boundary, before the prompt is read. It covers baselines, instruction files, MCP schemas, subagents and cache writes.
Codex Starts Encrypting Sub-Agent Prompts
Codex now encrypts the prompts it passes to sub-agents, according to a GitHub item. This is relevant to developers who work with Codex sub-agents.
pingdotgg/t3code GitHub repository listing
A GitHub repository from pingdotgg. No official description is available yet, so its exact purpose is not confirmed.
Image: GitHub
awesome-llm-apps: 100+ Agents and RAG Apps
A free, open-source collection of 100+ AI agents, agent skills and RAG apps, tagged for agents, LLMs, Python and RAG. Useful for developers looking for examples.
Image: GitHub
Clawk: Disposable Linux VMs for Coding Agents
Clawk gives coding agents a disposable Linux VM instead of the developer's laptop. It is tagged for Claude Code and Codex, among other agents.
Image: GitHub
Picked from public sources (GitHub, Hacker News, DEV, Lobsters, Reddit, Product Hunt) by our daily scout. Each summary uses only what the source itself says; GitHub stars are as checked on the date shown.