Why build it
Tracking competitors by hand has a failure mode everyone knows: you check obsessively for two weeks, then stop, and three months later you discover a rival shipped the feature you were planning. I wanted a system that watches a market every day without me, and that I could actually trust when it told me something happened.
The second half is the hard part. An LLM agent that reads the web will confidently report launches that never happened. Most agent demos ignore this. Rhizome Logic is designed around it.
How it works
Three scheduled agents run on the OpenClaw framework, sharing a file-based memory. One collects, searching each tracked competitor in parallel every day. One delivers, sending a digest to Telegram. One synthesizes weekly, reading across competitors to find patterns no single day shows.
Keeping an agent honest
The agent proposes, I approve.
Every fact the collector writes carries a source citation and a confidence score of high, medium, or low. Anything it could not verify gets flagged as unverified rather than silently included. And nothing enters the permanent record until a human has reviewed it. That gate costs me a few minutes a day and is the reason the record stays trustworthy: the agent proposes, I approve.
Memory is plain markdown files per competitor: one file each for pricing, hiring, product, and signals, plus dated snapshots of the raw pages. No vector database, no framework magic. When I want to know why the system believes something, I open the file and read the citation. Boring storage is a feature in a system whose whole job is being believable.
What it costs
About $23 a month in API usage for daily coverage of a market. The scheduled agents do the heavy reading; the human gate keeps the output worth reading. It has run unattended since late 2025, and the weekly synthesis has caught pricing moves and hiring shifts I would have missed on my own.
What I took from it
Agent reliability is a design problem before it is a model problem. The interesting engineering here was not the prompts, it was deciding what the agent is not allowed to do: write to the record without a citation, claim confidence it cannot back, or skip the human gate. Constraints made it useful.