Agent systems · Playbooks · Work in progress
Working AI systems from my own setup, each written up as a playbook you can copy: the problem, how it works, the guardrails, and what broke along the way.
Playbook 01 · Pattern: Triage Live
My agent asks a lot of questions that don’t matter. A small Jev model answers the low-risk ones and screens risky shell commands, so only real decisions reach my phone. Questions about purchases, money, messages, deletes, credentials or production skip the model and always come to me.
Playbook 02 · Pattern: Converge Live
School, sports and Scouts each use a different app. Every two hours an agent reads them all, read-only, and sends one combined update. If nothing changed, it stays quiet. If a source fails to load, it says so instead of pretending it was empty.
Playbook 04 · Pattern: Lanes Live
I run several agents. When more than one watched the same inbox and parent channels, I got duplicate checks and the risk of one agent acting on another’s job. Now every topic has exactly one owner, and a handoff isn’t done until the old job is confirmed off.
Hermes Agent
GmailPlaybook 05 · Pattern: Fan-out Live
A lead agent in Codex keeps the judgment and the approvals. It hands clearly scoped tasks to GPT-6 Luna at maximum reasoning, then checks the evidence before accepting. It’s stable, with a public setup kit and a video walkthrough.
Playbook 08 · Pattern: Meter Live
I wanted to know which models actually do my work, without keeping a log of everything I’ve asked them. Nightly collectors pull usage metadata from Codex and Claude into a local database. Prompts, responses and tool contents are never stored.
03 · Gate Testing
A Jev screen labels new mail for phishing and prompt injection; elevated risk is a veto for every other mail automation.
06 · Policy fork Testing
Narrow, reversible fixes inside a written policy; everything else becomes a read-only diagnosis for me.
05 · Fan-out Testing
Claude Haiku 5.5 takes bounded text tasks from a coordinating Claude model, which checks the exact model on every call.
05 · Fan-out Testing
GPT-6.1 Terra handles computer use: browser and desktop tasks. In a three-run test it wasn’t reliably faster yet.
Retired Capturing short-lived secrets locally so the model never saw them.
The value still passed through the messaging service, so it wasn’t the protection it looked like. Not a method I recommend.
Retired Earlier agent gateways, replaced by Hermes.
Retired names linger in old files and notes. They must never be mistaken for live services.
Retired AI-assisted matching between two sets of equipment records.
Coverage is not accuracy. Some rows simply lacked the data to match.
Retired Moving monitoring between agents by handoff note, back and forth.
One owner per topic, and a handoff isn’t done until the old job is confirmed off. See Playbook 04.
A calendar brief reported “no events” when its sign-in had expired.
Unavailable is not empty. A source that can’t be read gets named, never treated as quiet.
Building every component of a receipt-intake redesign before any record went all the way through.
Prove one record end to end first, then build out.
Trusting a worker’s “done” for a change to an outside system.
For external writes, check the target itself.
Comparing two prompts while shared memory leaked into both runs.
Test isolation fails quietly, and two trials don’t support a general claim.
Building a website traffic “AI report” whose first version turned out to be fixed text.
Check that the data can answer the question before building the report.
OpenRouter
Python
iMessage groups
BAND
ClassDojo