CASE STUDY · PRIVATE

organism

The private operation substrate: a governed agentic OS of 25 system pools that runs the whole single-operator business. The subject of the first essay.

Status ruling private, gate under NDA private repository, walkthrough on request

Problem

Prompt-era automation could not operationalize: agents had chat, not utilities. The operation needed one integration layer where agents get real tools, memory, and permissions under governance.

Build

  • Runs on a $20/yr VPS: no headless Chromium in the serving path, so every utility had to be a native binary or it did not exist.
  • A governance cortex: every architectural decision and regression guard indexed and enforced by a linter, so agents cannot drift from the rules by accident.
  • A swarm substrate: a task state machine, employee contracts, and a communication bus; since the October reorganization, one scheduler for the whole operation.

Outcome

  • Runs the whole single-operator operation: 25 governed pools across cortex, swarm, foundations, and ventures, each with declared service contracts, on a single $20/yr host.
  • The 476 telemetry failure (92 cron jobs declared, zero recorded executions) is registered as a reconciled figure; the artifact on this page reconstructs that recorded state, with the remediation that followed. The runtime it measured was decommissioned in the October reorganization.
Artifact: cron fleet, recorded state (September runtime)
$ hermes cron status --all
gateway: DOWN (hermes-active)
jobs registered: 92
recorded executions: 0
execution record: executions.db (0 rows)

# reconstruction of the recorded state, 2026-09-22.
# the live command returns current counts.

The failure and its remediation travel together: the worklog substrate, the fail-closed smoke gate, and the failure protocol are what now detect and block this state. The runtime that produced this record was replaced by the single-scheduler organization; reconciliation R15 records the retirement.

Provenance

Reproduce

The substrate is private, so the walkthrough is the reproduction path: `arch stats` re-derives the governance counts on the operator's machine, and Essay 1 prints the same table with the command named beside every row. The telemetry reconciliation stands on the September execution record: it holds zero rows while the declared job count is 92, and the claim that contradicted it was a hardcoded default, recorded as reconciliation R6.

From the documentation

  • The documented automation layer this portfolio runs on: a cron-fleet dispatcher, a registry, and an operator state store, described-not-linked because the repository holds the operator working state.

    walkthrough knowledge, repository private

  • The gateway-down failure state published on this study is reconstructed from the recorded incident, labelled reconstructed in the artefact itself.

    recorded incident, labelled reconstructed