Jay Margalus
◆ Agents · Resources

Tools and reading.

One real tool per job, each with the reason I use it, which is the part that transfers when the tool changes. Then the reading, grouped by what each piece changed for me.

Tools

The one I'd start with

  • Claude CodeAnthropic · terminal, IDE, desktopThe agent I run everything through. It works in your files on your machine, asks before it acts, and takes rules, skills, hooks, and memory. One real tool, learned deeply. Projects 1 and 2.
  • CLAUDE.md and memoryClaude Code docsThe rules file it reads every session, and the memory it carries between them. This is where your judgment gets written down.
  • Skills · HooksClaude Code docsSkills package a job you'll do again. Hooks are scripts that run around the agent and can stop it. A rule is a request. A hook is a fence.
  • Headless modeClaude Code docs · claude -pOne prompt in, one result out, no chat. The moment a job can be a sentence, it can be scheduled. Project 3.
  • Claude plans · ConsoleAnthropicFaculty: a Pro plan starts you, Max if you'll run Projects 3 to 5. Students in my courses get access through my Console organization; the invite comes by email. The desktop app doesn't take Console logins, so students use the terminal.

Reaching the world

  • Model Context Protocolopen standardThe plug standard for handing an agent a tool: mail, calendar, browser, database, your own script. Project 4.
  • PlaywrightMicrosoft · browser automationA real browser the agent can drive, and the way I check that a page looks right before anyone sees it. Ships an MCP server.
  • launchd and cronApple · UnixThe schedulers. Boring on purpose. Everything that runs for me without me is a launchd job or a cron line, listed in one manifest.
  • Tailscaleprivate networkHow my laptop reaches the always-on Mac mini from anywhere without opening anything to the internet. Set up once, forget it.
  • An always-on machineold Mac mini · Raspberry Pi · small VPSProject 3 needs a computer that's awake at 6am. Mine is a four-year-old Mac mini. A used one is a few hundred dollars and it's the best money in this whole list.

Local and private

  • Ollamalocal modelsOne installer, then ollama pull a model that fits your memory. Runs on the laptop, on a plane, with student data that can't leave. Project 5.
  • Open WebUIself-hosted chatA chat window over Ollama in your browser, on your machine, no account. The face on my router.
  • whisper.cpplocal transcriptionSpeech to text with no network. I dictate into the terminal with it and turn talks into text to read.
  • Obsidianplain-text notesWhere my notes live as Markdown files, which means the agent can read and maintain them. The wiki it keeps for me follows Karpathy's pattern below.

Also exists

  • FigmaprototypingThe earlier "one real tool" in my classes, for going from sketch to a prototype on a phone. The agent is the same bet one rung up.
  • Codex · Goose · Cursorother agentsFine tools, and I route some coding work to Codex. Learn one first. The second one takes an afternoon once you own the first.
  • Not recommendedsubscription proxiesTools that route your Claude subscription login through a third-party router or proxy. Anthropic's terms don't allow it, and the ones I tested broke. Use the official tool, or pay per token with an API key.

Reading that got me here

Pedagogy, which came first

  • MindstormsSeymour Papert · 1980 · free PDF from the Media LabChildren learn by building with the computer, not by being taught about it. Every project on this page is Papert with a terminal.
  • Situating ConstructionismPapert & Harel · 1991The three tests for a learning artifact: personally meaningful, publicly shareable, connected to a community that cares. An agent you keep passes all three.
  • Anthropology and/as EducationTim Ingold · 2018, reissued 2025 as Old Ways, New PeopleThe teacher's job is making the context, not transmitting the content. This page is a context. The lecture version wouldn't work.
  • Entrepreneurial Education: A Path to EmployabilityTimothy Hor · AACSB Insights · March 2026When polished output is cheap, what employers pay for is judgment and ownership. Grade the decision trail, not the deliverable. That's the faculty section above.

Agents and harnesses

  • Building Effective AgentsAnthropic · Dec 2024Simple, composable patterns beat frameworks. Add complexity only when it measurably helps. I reread it whenever my setup starts to feel clever.
  • Harness Design for Long-Running AppsAnthropic · March 2026Separate the builder from the evaluator. This is why a second agent checks every deploy I make, and why Project 4 has a verifier.
  • AI Fluency: Framework and FoundationsAnthropic · free courseDelegation, description, discernment, diligence. The vocabulary I use with alumni. Idea 11 on this page is discernment.
  • LLM WikiAndrej Karpathy · 2026Let the model maintain a linked wiki between you and your raw sources. My vault runs this way. Also his autoresearch: hundreds of cheap overnight experiments, keep what passes a verifier.
  • Eight levels of AI adoptionSteve Yegge, via The Pragmatic Engineer · 2026A ladder from autocomplete to running a fleet of agents. Useful for placing yourself honestly. Most of this page lives at levels four to six.
  • The Evolution of the Agent HarnessDan McAteer · Latent Space · Aug 2026As models absorb what harnesses used to do, what's left is managing human attention: permissions, identity, trust, legibility. ATTENTION.md exists because of this piece.
  • wayfinderMatt Pocock · skills repo · 2026Plan work too big for one session as decision tickets, each one a question whose answer is a decision. How I run overnight work on the Mac mini.
  • Introducing dotsOpenAI · Sept 2026Always-on agents with their own computer. I built my own version the week after, on my own machine, because mine can read my university mail and theirs can't.

Local models and routing

  • Running Local Models Is Good NowVicki Boykis · June 2026About three quarters of frontier quality on refactors, tests, and proofreading, on a laptop. The post that made me take Project 5 seriously.
  • Ahmad Osman on why local AI is catching upLatent Space · June 2026A model is only one part of the system. The local agent that failed was missing search, not intelligence. Four to eight months of lag and closing.
  • Why model routing is in demandLatent Space, with Glean · 2026A cheap model gathers context and decides "enough or escalate" before the expensive one is called. Half the traffic never reaches the frontier. The shape of my router.
  • How we rebuilt provider routingOpenRouter · Oct 2026Route on measured cost, speed, and reliability, not list price. I borrowed the scorer the week it came out. It's the second screenshot in Project 5.

Books on the table

  • Co-IntelligenceEthan Mollick · 2024The book I assign first in ENT 1000. Treat the model as a colleague with a jagged frontier, and always invite it to the table.
  • The Minimalist EntrepreneurSahil Lavingia · 2021The second book in the course. Build for a community you're in, charge from day one, stay small on purpose. An agent that does one tedious job for one real business is a minimalist company.
  • How Buildings LearnStewart Brand · 1994Buildings get remade by the people in them. So do harnesses. The "where it breaks" section above is this book applied to a Mac mini.
  • Working in PublicNadia Eghbal · 2020Maintenance is the real work and it's invisible. Read before you schedule anything you'll have to keep alive.
  • The Infinity MachineSebastian Mallaby · 2026The DeepMind story, which is also the story of why the model is the part you'll never control. Handed out at the alumni workshop.
  • A Turbulent AI Era and Critical Choices to MakeBill Gates · Gates Notes · Aug 2026Even the good version of this transition is rough. Faculty asked me for one general-audience piece to assign; this is it.
  • Why Is Digital Freedom Making Us Exhausted and Sad?Matt Elmore on Byung-Chul Han · The New Atlantis · 2026The counterweight. An always-on agent that brings things up first can be a tool for attention or another way to be seduced by the screen. ATTENTION.md is my answer; Han is the question.

Mine, for the record

Links checked Nov 1 2026. If one rots, tell me.