Agent systems
Built on my own time
Designed and built alone · my own server
A second brain with agents, running unattended since April 2026
More projects than time, and a brain that starts more than it finishes. So I built one for myself: a knowledge base of eight hundred notes, wired to agents that run on schedule. Six are live today. The design problem was never the automation. It was probabilistic output: what an agent may claim, where a human has to be, and what happens when it is wrong.
- 6agents live, unattended since April 2026
- 809notes in the knowledge base
- 1verb per agent, one directory it may write to
01
One agent, one verb
Eight agents reached production and six run today: an external watch that scores what it reads and keeps only the strongest items, a commitment tracker, a voice-to-note capture, a weekly rewrite of the system's own context file, a morning brief, and a content pipeline. Three more were specified and deliberately never built: low frequency, high judgement, moves files. That is the worst possible profile for an unattended agent.
Each agent owns one directory and writes nowhere else. Two messages a day, maximum. Alerts only above a score of seven. An agent that talks too much gets muted by its owner, which is the real failure mode. And the last step stays human: nothing reaches a public channel without me.
Daily watch · 5 items kept out of 212
source first, always: the link comes before the agent's summary
Six lines: what it is, why it matters to a project, what is uncertain.
Score 8. Flagged: the figure in paragraph two has no source.
The decision surface: the source link first, the agent's summary second, three buttons, no free text. Pressing a button is the record.
02
Provenance beats fluency
The first briefs read well and contained claims I could not source. Every factual claim now has to trace to a named file before it can be written, unsourced figures are frozen in a list the agent may not use, and anything decisive but uncertain is flagged on the card instead of buried.
03
What broke, and what it taught me
Silent success is the enemy. One agent produced empty output for fifteen days while reporting nothing. Three others were found dead by an audit, one after three weeks. The fixes were parsing and quotas, but the lesson was supervision: nothing had pinged when an agent went quiet. A health check that shouts when an agent stops is the first thing I would build anywhere else.