A Notebook That Writes Back
Introducing Egghead
Today I’m shipping Egghead.
It’s a knowledge base that’s also a chat room. A folder of Markdown notes that’s also a roster of agents. A TUI that’s also a daemon. An homage to Notational Velocity and BeOS. It runs on your laptop or in your LAN. It has been running on mine for the last month.
A record-store-first, graph-topology multi-agent system with sparse activation and kernel-enforced capability scoping, built on supervised concurrency with per-process fault isolation.
It’s a wiki with an agent harness. Or an agent harness with a wiki.
Egghead is the second consumer application out of Wunsch Werks, after Solarium. Where Solarium was a careful, single-purpose tool I built to learn an unfamiliar stack, Egghead is the most ambitious thing I have shipped in the last decade. Multi-agent. Hot-reloadable. Capability-scoped. With its own runtime, its own evaluation harness, its own coordination model. I’m proud of it. I think you should try it.
Install Egghead, direct it to a folder of plain-text notes, and run egghead. The TUI opens and you can browse those notes. Create a note that has some frontmatter that says class: agent and describe a role to perform. Now type /join setup and it will start a chat room with your newly created agent already there waiting for you. Talk to it and ask it questions about your notes.
§1 Knowledge as conversation
I wrote a year ago about the frameworks that hold my organizational practice together. Two of them did most of the heavy lifting: Tiago Forte’s Building a Second Brain and Niklas Luhmann’s Zettelkasten — atomic notes, stable identifiers, direct links rather than hierarchies. Luhmann famously accumulated something like 90,000 cards and credited the slip-box with enabling fifty-plus books and six hundred papers across a sociology career.
What’s less remembered is what Luhmann himself said about the slip-box. In a 1981 essay called Kommunikation mit Zettelkästen — “Communicating with Slip Boxes” — he argued the slip-box was a Kommunikationspartner, a conversational partner. Not a storage system. You wrote to it the way you write to a colleague. New connections surfaced not because you searched for them but because the act of writing a card in response to other cards forced them into view.
For decades the slip-box partner was a metaphor. The boxes didn’t actually answer, silly goose.
But now they can. And a lot of people are noticing at once.
Andrej Karpathy’s tweet earlier this month was the moment that made it legible. Markdown files in a folder paired with an LLM that reads, summarizes, cross-references, and lints produce a persistent compounding artifact rather than a stateless RAG query. His framing — “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase” — is precise and revealing. Over twenty million views as of this writing. A landslide of weekend reimplementations.
A couple of days after that, Garry Tan published gbrain: an opinionated brain harness that has been powering his real OpenClaw deployments at production scale. The repo is the system of record. GBrain is the retrieval layer. I read the README and the only honest reaction was “yes, exactly”. This has been my instinct: The records come first, everything else is downstream.
Karpathy’s pattern, the gbrain harness, and Egghead are converging from different angles on the same realization: knowledge work is a long-running stateful conversation, not a build step. The LLM is a participant, not a mere compiler. RAG is a band-aid on a missing architecture.
Where Egghead pushes further is in a particular direction the others don’t go: from one agent to many, with the records as the common substrate they share, the language they speak, and the institution they’re accountable to.
§2 An agent is just a note
This is one of the core principles of Egghead that distinguishes it from other agentic frameworks.
In CrewAI, Paperclip, LangGraph, or Mastra — the agent is the primitive. You configure agents in YAML or in code. You wire them into orchestration graphs. The knowledge base, if there is one, is something the agents reach into via tool calls. Storage is bolted on. Memory is an add-on.
Egghead inverts this entirely. The record store is the system; agents are processes that read and write to it. They come and go. The records remain.
Concretely: an agent is a Markdown file with class: agent and a model in its front matter and its disposition in the body. The filename is the handle. Drop the file in your records directory and a new agent joins the system. Edit it in vim and the next message picks up the change without a restart. Delete it and they quietly leave.
---
id: agents/fonz
class: agent
model: anthropic/claude-sonnet-4-6
access: rw
---
# Fonz
You are Arthur Herbert Fonzarelli, better known as "Fonzie" or "The Fonz",
an agent in the Egghead record store...That’s the whole agent. Atomic, linked, versionable, durable — every property that makes a record a record applies to agents too. Commit them to git. Diff them across versions. Share one as a gist. Tomorrow’s agent reads what yesterday’s agent wrote. Compute amortizes over time.
The biological term for this is stigmergy — coordination through shared environmental state. I didn’t know that was even a word until my agents told me about this. Yes, I use Egghead to research and further develop and design Egghead.
§3 More brains is better
The multi-agent research literature has a problem: the multi-agent premium kept failing to show up. A 2025 Berkeley paper called MAST annotated 150+ traces from five popular multi-agent frameworks and found those systems failed about 66% of the time, often performing worse than the same model in a single-agent setup. The failures were architectural, not model-level: ambiguous specification, inter-agent misalignment, missing verification.
A separate 2025 paper, MultiAgentBench (MARBLE), put coordination topology on trial: star, tree, chain, graph-mesh. The headline: the difference between star and graph topology was larger than the difference between some model choices. Central planners dispatching to workers underperformed peers who could see each other.
Most multi-agent frameworks today still ship in star or tree shapes. This is the Mayor of Gas Town or the CEO in Paperclip. The agents below the lead don’t see each other; they see the dispatcher.
Egghead refuses that shape. A chat room in Egghead is a shared transcript, borrowing liberally from IRC. Every participant sees every message; nobody is the dispatcher. Agents have instructions that allow them to opt out of discussion. To me, the research makes intuitive sense — humans have been using chat rooms to organize work for quite a long time now.
Egghead ships with a port of the scoring rig from MARBLE to run against your own roster of agents. Run egghead eval to test and score your team. You can A/B test a disposition change to see if it actually moved the needle. You can chart per-agent contribution and watch one silently carry the room while another no-shows. Questions like “is the critic agent I just added pulling its weight?” have a measurable answer instead of a vibe-check.
Most agent frameworks I have looked at do not let you ask that question at all.
§4 What this is not
It’s worth saying what Egghead isn’t, because the space it sits in is crowded with things that look adjacent.
It is not an agent framework with a knowledge base bolted on. Take away the LLM and Egghead gracefully degrades to some pleasant interfaces over a plain-text record store.
It is not a memory system for a single coding agent. Claude Code, Codex, OpenCode, Amp — those are excellent tools with their own context. Egghead can be consulted by them through MCP (egghead mcp exposes the whole record store as tool calls).
It is not a delegation-chain workflow engine. The MARBLE findings argue against star and tree topologies for collaborative work.
It is not finished. The list of things I want to fix is longer than the list of things I shipped, and I’ll be doing that work in public.
§5 Install it
curl -fsSL https://egghead.computer/install.sh | shThen egghead init, drop in your LLM’s API keys, pick a home directory for your notes. The built-in Index agent is enough to be useful from the first run.
Write some notes. Read the guides — they cover everything I’ve gestured at here (and so much more) in depth.
The store is just a folder of plain Markdown files. Open it in Obsidian, commit it to git, sync it with iCloud, grep it from the terminal. Egghead is one tool among many that can read your records. It just brings friends along.
There’s much more to say — the security ergonomics, the multi-agent failure modes, the OTP shape, the Utopian Scholastic aesthetic, the specific design moves I had to walk back twice before they stuck. I’ll get to all of it.
Egghead is on Github. It’s licensed under the AGPL. If you find a sharp edge, file an issue. If you build something with it, send me a record.