← all work · personal project
Exocortex, a multi-agent AI system I run daily
A plain-file knowledge base worked by several AI agents, with task routing, guardrails and human approval on anything that leaves the machine.
- Problem
- One person, many knowledge workflows (notes, tasks, research, planning). I wanted real help from AI without vendor lock-in and without my data leaking.
- My role
- Designer, analyst and daily operator.
- What I did
- Plain Markdown files as the single store; several AI coding agents working the same files; a task queue where every task is a written packet with a route and a done-test; guards on every write.
- Result
- In daily use since April 2026. Every task has an owner, a route and a check, and the guards have stopped real mistakes before they landed.
Follow one task through the system
Pick a task, then step through it.
- A written briefa goal, the facts, a done-test and a route Compare five note apps on privacy. Done when every claim has a source. Tidy the task queue and publish the result.
- Routed by its shapebulk reading goes to a cheap model, decisions to a strong one Reading five websites, so a cheap, fast model. Editing files, so a strong model.
- An agent does the workon the same plain files I use Reads the sites and writes a table with a link for every claim. Tidies the queue, then tries to push it online.
- Guards check every writeno secret paths, no pushing, no changes to its own rules Every write allowed. The secret scan is clean. Blocked. Agents may not push. The task stops here and waits for me.
- I approve what leaves the machineanything sent or published in my name Nothing leaves the machine, so no approval is needed.
- Closed with a reportwhat was done, how to check it, what it cost The done-test passed, the report is filed, the cost is logged.
Why plain files
Every note, rule and task is a Markdown file in one git repository. Any tool can read it, nothing is locked inside an app, and every change has a history. The AI agents work on the same files I do.
Tasks as packets
A task is a small written brief with a goal, the facts it needs, a done-test and a route (which model, how much effort). An agent claims it, does it, and closes it with a report. This is requirements work applied to AI. An unclear task produces unclear output, so the brief is where quality starts.
Routing by task shape
Cheap, fast models do bulk reading and search; stronger models do design, decisions and code. The route is chosen by what the task looks like, and cost is tracked per task, so spend is a decision, not a surprise.
Guardrails and the human in the loop
Agents cannot push, cannot read secret paths, and cannot change their own rules without my approval at a real keyboard. Commits are scanned for secrets. Text an agent reads from the web or a message is treated as data, never as instructions. Anything sent in my name waits for my approval.
What a team can take from it
Write the brief before the prompt. Route work by its shape. Check before you trust. Guard the write path. Keep a human approval where the action leaves the building.