Automate the digital, surface the physical — HITL and HOOTL for ensembly

Automate the digital. Surface the physical. Wait only for permission. Make the truth playable.

I am not playing a game

There is no recreational play loop here—no dopamine farm, no strategy for its own sake. The interface may borrow game affordances later: rhythm, spatial layout, a map of what matters. That is taste, not addiction. Joining is a deliberate act. Human-in-the-loop (HITL) is an optional verification and steering surface. I step in when I want to confirm that payments, obligations, queues, and systems are progressing as expected—or when a gate demands my body or my explicit approve/deny.

Fast agent stacks are not the enemy. Grok Bot is smooth; I use it. The question is what a life OS must own when the model session ends: the graph of what depends on what, the critical path through it, and the law that auth and physical gates never self-approve.

While I am out: HOOTL

Human-out-of-the-loop (HOOTL) is the default working mode. A helper swarm operates against the current state of life—curating, prioritising, balancing, clearing digital thrash so scarce attention stays reserved for two irreducible classes: body-world pickups (only a physical presence can complete them) and authorization gates (real risk, irreversibility, explicit permission).

Prioritisation is not a flat to-do list. Urgency and importance form a directed graph of strongly connected, sequentially dependent clusters. From that graph the runtime derives the critical sequential path—the chain whose timely resolution determines overall health. PERT and Monte Carlo are tools I may use to quantify uncertainty and surface bottlenecks. They are not the product story. The story is: the swarm works the path the kernel names, not whatever is loudest in the inbox.

Two layers, one truth

The architecture splits cleanly. The kernel / runtime owns life-state, the dependency graph, critical-path calculation, HITL escalation, triggers, and the message bus. It is the thin structural source of truth. Intelligence is rented or open—I do not grow a new organism inside the kernel each day.

Workers sit underneath: multi-agent helpers that execute digital chores, report status, and return verified artifacts. They coordinate through the graph and the bus, not through ad-hoc chatter. Game clients, Unity scenes, GPS beacons can sit on the same protocol later. They are not the core. The core is coordination, invoke/trigger, concurrent state, and typed messaging.

That split mirrors what the memory essays argued from another angle: pulse instead of dump, archive admissions instead of log completeness. Here the kernel pulses priorities; workers inject surgical results; HITL surfaces only what still needs a human.

What remains hard

Stripping optional game chrome leaves four engineering problems that actually determine quality:

1. CoordinationConcurrent agents on shared life-state without conflict or thrash. The dependency graph is the coordination substrate—claim work through it, not around it.
2. Invoke / triggerTime-sensitive activation from life-state changes, schedules, or external events. Declarative triggers beat scattered conditionals.
3. State updatesConsistent, auditable mutation of life-state and graph. One narrow interface; ordered or transactional where it matters.
4. Message passingTyped, reliable communication between kernel and workers—manual commands, automatic triggers, and status reports on the same bus.

How I measure outcomes

I will not map biology one-to-one. Evolution is slow; my operator has one life. Generational gain comes from harness design and evals, not from pretending every day is a new species. Evolution-shaped loops are reserved for multi-session blanks that cannot close in one tick—not for clearing email.

Verified artifacts answer three questions:

correct?

Does the result match the job spec and pass HITL auth when required?

effective?

Did it save future energy and effort—shorter critical path, less thrash tomorrow?

efficient?

Did it spend less time and cost for this step—probe budget, wall clock, tokens?

That triad is closer to active design-of-experiments and robust decision-making under uncertainty than to open-ended Darwinian machine evolution. The swarm stays structural; the models I call are interchangeable renters.

Closing

The EEaaS to agents essay insisted that the cost of the next useful observation must fall while decision quality rises. Pulse instead of dump made that operational for memory. Archive is not memory asked what may enter the harness at all. HITL and HOOTL ask who runs the day: the swarm on the critical path by default, me at the gates when judgment cannot be delegated.

Automate the digital. Surface the physical. Wait only for permission. The human stays the scarce resource; the kernel keeps the truth thin and playable.