Welcome back to Agentic Coding Weekly. Here are the updates on agentic coding tools, models, and workflows for the week of May 10-16, 2026.
Executive Summary:
Claude Code adds
/goalcommand for autonomous goal-directed work. Haiku judges completion after each turn.Agent View ships for managing multiple Claude Code sessions from one screen via
claude agents.Claude Code weekly limits up 50% through July 13. Subscription changes coming June 15 for programmatic usage.
x.ai launches Grok Build CLI in early beta. SuperGrok Heavy only ($300/month).
Codex now works from ChatGPT mobile app if you have the desktop app running.
Worth reading: Why deliberately using non-SOTA models might give you better code quality, the math on why AI that doubles your output might quadruple your maintenance costs, and why senior engineers often struggle to communicate their expertise.
Meanwhile, another week, another new coding agent. Checkout zerostack, a minimal coding agent written in Rust with only ~12MB RAM footprint when working.
1. Tooling and Model Updates
/goal Command in Claude Code
After Codex two weeks ago, Claude Code added the /goal slash command for goal-directed autonomous work. Instead of stopping after each response, Claude keeps working toward a concrete objective.
Run /goal with a verifiable objective. For example, /goal all tests in test/auth pass and the lint step is clean.
How Claude Code decides the goal has been achieved? After Claude finishes responding, the goal and the conversation is sent to Haiku model which returns a yes or no decision and a short reason. Yes stops the loop, no sends the reason back to Claude and asked to keep working.
Agent View
New way to control and manage multiple Claude Code sessions from a single screen by running claude agents. GUI-based IDEs and ADEs like Zed, Cursor, Codex, and Claude desktop apps have had this for a while. Now it's time for TUIs to get it. I'm holding on to my tmux panes and windows though.
Claude Code Limits and Subscription Changes
Weekly limits are increasing by 50%, currently through July 13.
Starting June 15, 2026, Claude subscription covers only interactive use (TUI) of Claude Code. Users get a separate monthly credit equal to their subscription amount for programmatic usage like claude -p, third-party apps using the Claude Agent SDK, or tools like openclaw.
You have to manually claim the credits once, then they refresh automatically each cycle.
Quick Updates
Codex is now available in the ChatGPT mobile app. Requires Codex desktop app running on one of your machines. Not the Codex CLI, the Codex app.
x.ai launched their own proprietary CLI coding agent, Grok Build CLI. Early beta, available only to SuperGrok Heavy subscribers ($300/month plan).
2. Community Picks
Why Senior Developers Fail to Communicate Their Expertise
I like how it explains the two competing needs in a company: the need to keep existing services stable and the need for moving fast with new services and products and how these two affects developers.
The solution offered by the author though I think doesn't really work that well in practice. PoC that was never meant to run in production running in production is a tale as old as time.
Running Local Models on an M4 with 24GB Memory
Shows how to run Qwen 3.5-9B (4b quant) locally with pi coding agent along with a couple of example coding tasks. Runs with around 40 tokens per second and has a 128k context window. Obviously doesn't match SOTA models, but works for small focused tasks like git conflict resolution.
Found an interesting insight is buried deep in the article: you might deliberately want non-SOTA models for coding.
Since these smaller models require more hand-holding and guidance, you have to work through the problem yourself. You end up with a better understanding of the code and a stronger mental model of your codebase. With SOTA models, it's easy to offload all cognitive effort. With smaller models, you're making all the architecture decisions, data organization choices, everything. You'll likely end up with better code quality.
Read the post and the discussion on HN.
You Need AI That Reduces Maintenance Costs
If you want a productive team, you have to focus on their maintenance costs....
The math only works if the LLM decreases your maintenance costs, and by exactly the inverse of the rate it adds code. If you double your output and your cost of maintaining that output, two times two means you’ve quadrupled your maintenance costs. ...
Instead, you have to invert your productivity. If you’re producing twice as much code, you need code that costs half as much to maintain. Three times as much code, one third the maintenance.
Read the post and the discussion on HN.
That’s it for this week. I write this weekly on Mondays. If this was useful, subscribe below:
