Reading Room

You are thinking about how AI agents change team and project structure, and this piece gives you a sharp, specific lens for that instead of vague hype.

The author opens with a true story: they watched four people struggle to connect a laptop to a screen, and when someone asked them to help, they said no, because adding more people would not solve it any faster. That is a direct nod to Brooks's Law, named for computer scientist Frederick P. Brooks Jr., who wrote it in his 1975 book The Mythical Man-Month, based on his own experience running IBM's OS/360 software project: "Adding manpower to a late software project makes it later."

Their question is whether that law still holds now that AI coding agents (AI tools that can take several steps on their own, writing and running code without a person typing each line) are doing real work. At first it looks like no: an agent is ready instantly, does not need onboarding, and does not take coffee breaks, so the old people problems, slow ramp-up, meetings, office politics, seem to vanish.

But they argue the law just changed shape instead of disappearing. The "Baby Problem" still applies: nine agents cannot produce one result in a ninth of the time if the steps have to happen in order, the same way nine people cannot deliver a baby in one month. The management overhead has not gone away either, it has moved. Instead of managing people, you are now managing agents, and a person still has to decide what "done" means, check the code, catch the places where the model made something up, and fit the finished pieces together. That someone, they write, is you.

The sharpest risk they name: hand nine agents the same codebase with no clear plan, and they will rewrite each other's work in a fast, repeating loop, undoing progress instead of adding to it. Their proposed update to Brooks's Law for this moment: "Adding agents to a project with unclear requirements gets you the wrong answer faster." The lesson for anyone handing off work to several agents at once is the same one Brooks gave teams fifty years ago: the bottleneck was never headcount, it is clarity, and more hands, human or AI, just make an unclear plan fail faster.

It is a fast way to see which AI tools developers are actually adopting, not just talking about, matched against what you have already starred yourself.

The post opens with scale: GitHub (the site that hosts shared code) holds 450 million repositories in total, then narrows to 20 AI-related ones worth knowing, ranked by how many developers have starred them. A star is GitHub's version of a public bookmark: a developer stars a project to mark it useful or worth tracking, so the count is a rough measure of real attention, not hype.

The biggest group is AI coding agents (tools that write and run code on their own): OpenClaw leads with 278,000 stars, ahead of Opencode at 118,000, Claude Code at 75,000, Superpowers at 73,000, and Codex at 63,000.

A second group covers agent engineering and tools, the plumbing used to build agents rather than the agents themselves: Firecrawl at 89,000 stars, Context7 at 48,000, Scrapling at 25,000, Agent Browser at 19,000, and Symphony at 8,900.

A third group is AI infrastructure, the lower-level software that runs or serves models: Open WebUI at 126,000 stars, llama.cpp at 97,000, Daytona at 63,000, and Zeroclaw at 24,000.

The fourth group is learning material: Awesome LLM Apps at 100,000 stars, AI Agents for Beginners at 53,000, Prompt Engineering at 32,000, Hello Agents at 25,000, and Hermes Agent at 200,000. One final entry, System Prompts of AI Tools, is a curated collection sitting at 129,000 stars.

Read across the five groups, coding agents and learning guides draw the largest followings, while the infrastructure and plumbing tools, less visible but doing more of the actual work underneath, draw smaller but still substantial numbers. The post closes by asking readers which of these they are building with, treating the list as a conversation starter rather than a verdict on which tool wins.