Loom

Loom is my research assistant. I give it a question. Guarded helpers search and read the web, and Loom writes a report that cites its sources. Each new pass starts from what earlier passes found.

Why it exists

Research for a project rarely happens in one sitting. Sources come in over weeks. In my own words: "over time I collect sources, its not that I am researching something and I have all sources handy."

The value comes when old research feeds new research. As I wrote when I set out its vision: "I see this as a compounding effect."

Loom once held my software delivery work as well. That work moved to Atlas, and Loom kept one job. In my words: "this project is now a researcher".

A research pass

Each topic lives in a thread. A pass on a thread starts with a brief, and each pass runs only after I give it a go. First, Loom reads what earlier passes on the thread already found.

Then the work splits on purpose. One agent may only search the web. A separate reader with no tools reads the pages it finds, and it refuses to start if an Engram token is in its environment. A check rejects a report that breaks its rules. Findings from web searches go into Engram only after I say so.

Loom can also build an illustrated explainer from what it knows. It has built two so far.

Sources over time

Sources I find along the way go into a queue. I tag a saved item in Engram, and a worker reads each one closely when I start it. It marks each claim as new, already known, or in conflict with what Engram holds. New items go straight into Engram, marked as unreviewed, and new links arrive as proposals for me to decide.

A door for other agents

Agents in my other repos, on my own machine, can reach Loom through a plugin. They can list its threads and copy a cited report into their own repo. That half is live, and I proved it from the Atlas repo.

They can also ask Loom a new question, within limits I set, and some asks wait for my approval. That half is built and merged. It stays switched off until its launch proof runs again on the main branch. No agent has asked from another repo yet.

Where it goes next

The order is set. First, the launch proof runs again on the main branch. Then one real ask comes in from a second repo. Then comes the step that files what an ask found into Engram.

A refresh that I start by hand already works; a refresh that runs without me is not built yet. The source worker runs when I start it, and I will decide on a schedule after its first real batch.