Can I hand the planning and follow-up in my life to a system, and tell whether it actually helps? One orchestrator writes, memory comes in four kinds, every captured item needs a verbatim quote, a nightly job runs the planner in shadow mode, and a draft rubric tracks the effect over time.
Question
Can I hand the planning and follow-up in my life to a system, and tell whether it actually helps?
I think AI should take our busywork. It should make what we can already do far more efficient and leave us the thinking part. When I applied that to my own life, I started by tracking and checking the most basic things, and I add the next layer only once that level is in order.
The system runs on Hermes Agent, an open-source agent runtime.
Architecture
FIG. 1One path through the system. Seven layers on one Linux machine at home, with two trips through them: an item heard on the phone, and a piece of research a worker proposes. Dashed boxes do not act yet: the night planner runs in shadow mode, and no referee has passed its bar. Source: How I run my life sheet, Architecture (round 4 text); orchestrator charter (sources.md)
One agent is the only part that writes, and its charter starts with the rule that the system prepares and I decide. After every turn a fast model proposes the open items it heard, each with a word-for-word quote, and code drops any whose quote is not in the turn. A worker's "done" counts only with a file path, a URL, command output or an exact quote.
The night planner proposes where slipped tasks go in free calendar blocks, and a script checks every move. For now nothing gets written (shadow mode). The last audit check, the referee, is meant to catch the agent claiming work it did not do.
Measuring the effect
FIG. 2The rubric, first readings. Each reading with the direction it should move if the system helps. The first three should fall; the fourth watches the checks themselves. Source: How I run my life sheet, rubric table; session ledger and night planner records, queried 2026-11-05 (sources.md)
An audit in early October gave the first measurement: in my first six weeks on Hermes, 71% of my active session time went into building and running the system instead of my own work. One rater sorted the sessions, so the figure is rough, and those weeks were setup, when a high share is expected. I treat it as a starting point, to re-measure the same way.
It is one row of a working rubric where every row already has a data source:
Dimension
What is measured
Source
Time on the system
Share of active session time spent building or running the system
Session history, sorted by topic
Open loops on me
Open items waiting in the session ledger
Session ledger
Slipped commitments
Dated tasks past their time, counted each morning
Night planner
Proposals stopped by checks
Items thrown out by the quote check; planner moves rejected by the validator
Session ledger, planner results
Routines held
How many of my 4 basic routines met their own floor, against the week before
Sunday review
I count rejections because a check that never says no might not be looking.
What did not hold
FIG. 3Seven months, and one decision still open. The build ran for months before the first measurement. The hatched date is a decision that was due and has not been taken. Source: project.json timeline; How I run my life sheet, What did not hold
After the audit I cut the scheduled jobs; they have grown back, not all for this system. The quote check threw out few invented items; more noise came from the agent's own offer to go through items one by one, captured 18 times. On the day the rules were written, the agent labelled them approved before I had said yes. The planner is still in shadow, past the date when the decision on letting it write was due.
Open question
For anyone who measures what tools change in people's lives: most of what this system should change happens outside a chat session, where my audit could not see it. How would you measure that part without logging the rest of my life?