11 reads.
The clearest writing yet on why the model stopped being the bottleneck. Thariq's frame: the gap between what you asked for and what the work actually needs is where quality leaks out, and the real skill is surfacing those gaps before, during, and after you build. Worth reading twice.
The best long read on why a top-tier model isn't the point. The system around it is. Loops, memory, verifier sub-agents, and state files are what make each run leave the next one smarter, and the piece lays out the whole stack with cost-routing advice for when to reach for the expensive model versus a cheap one. Long, but it's the map most people are missing.
An Anthropic engineer on two things the top model changes: self-correction loops and memory. The sharpest takeaway is that the model shouldn't grade its own work. An independent verifier explores harder and recovers from dead ends where self-critique stalls at good enough. Short, concrete, and from someone who actually ran the experiments.
A plain-English tour of the Claude features hiding in plain sight: Projects, memory, extended thinking, scheduled tasks, prompt caching, custom roles. Nothing exotic, but the value is in the framing. Each one takes minutes to set up and pays off daily. Good to forward to anyone still treating Claude as a fancier search box.
The line most people skim: every agent in a workflow can run a different model. This is the proof. A 33-agent audit, cheap readers plus one smart synthesizer, ran in under five minutes and found a real buried error. Put the expensive model where the thinking happens and run the rest on the cheap one. That's the economics of parallel agents in one example.
A clean way to think about agent design: the model is deliberately thin, and intelligence gets pushed outward into memory, skills, and protocols that the harness composes at runtime. The useful question it hands you is where any new capability should live. Good conceptual scaffolding if agents still feel like a bag of tricks.
The canonical piece on dynamic workflows, from the Anthropic engineer who built them. Claude writes its own custom harness on the fly to beat the failure modes of one long context window: laziness, self-preference, goal drift. The example prompts alone are worth the read, and it names the reusable patterns (fan-out, adversarial verify, tournament) you'll see everywhere else. Start here.
Karpathy on a workflow more people should steal: point an LLM at a pile of raw sources and let it compile and maintain a markdown wiki you rarely touch by hand. Once it's big enough, you query it like a research assistant, no fancy RAG required. The best part is that your own explorations file back in, so the knowledge base compounds. One of the most-shared AI ideas of the year for a reason.
Karpathy's follow-up carries the sharper point: in the agent era you stop shipping code and start shipping the idea, kept deliberately vague, and the other person's agent builds it for their needs. It's a small reframe with big implications for how software gets distributed. The knowledge-base spec is the worked example.
A genuinely good step-by-step on running Claude Code 24/7 on an always-on Mac Mini so you can text it tasks from anywhere. The best detail is the one nobody mentions: there's no message queue, so a sleeping laptop drops everything, which is the whole case for dedicated hardware. Practical if you want a personal agent that never goes dark.
The reference for anyone configuring Claude Code seriously. It walks the whole control center: CLAUDE.md, path-scoped rules, and hooks, with the key nuance that instructions are suggestions but hooks are deterministic. Keep CLAUDE.md under 200 lines or adherence drops. This is the one to bookmark.