Executive summary · as of 2026-09-03

Higher Options AI Operations

Built from live reads of every host, the Cloudflare account, Healthchecks, GitHub, and gbrain. Source of truth: higher-options-agent-ops/docs (PR #6).

The shape of it

Higher Options is a 5–10 person recruiting firm (EA and offshore talent, plus the Aptive Index assessment) run by a non-technical founder. Since April 2026 the back office has been rebuilt as an AI-operated system. Claude does the planning, coding, review, and daily operations. A fleet of scheduled routines runs unattended. Every follow-up that needs a human lands in Asana.

Surface	Role	Notes
MacBook Air Claude Code + Cowork	Jason's seat. The daily brief, the autorun dispatcher, the build controller.	Sleeps, so not for unattended work.
Mac mini “CORA” Cowork routines	The Cora executive-assistant persona: inbox and Slack watchers, Galarm alarms.	M1 8 GB, OOM-prone. being retired onto Hermes
Hermes VPS Nous Hermes Agent v0.20.4	Always-on agent: Slack DM to Jason, nightly gbrain “dream”, missing-paperwork report, GitHub plan reviews, the GLM/Kimi second-opinion engine.	Sandboxed: own Cloudflare account, read-only Drive, drafts-only Gmail, scoped GHL bearer, egress allow-list.
Cloudflare Workers 17 in production	Serverless layer: MCP connectors, CTS, CORA bookkeeper, dashboards, watchdogs.	Deterministic and cheap. All stateful data lives here (D1, KV, R2, Vectorize).
claude.ai cloud routines	Daily-brief shadow run.	soak since 2026-08-29 Sandbox can't reach hc-ping.com; heartbeat relayed via Hermes.

Shared memory: gbrain (Supabase Postgres + pgvector), written by Claude Code, Cowork, and Hermes, searched before answering, consolidated nightly. System of record for anything needing Jason: Asana. Alert rail: Healthchecks.io dead-man checks (12) fanning out to SMS, Pushover Emergency, email, and Slack DM. Every automated message names the machine that sent it.

What runs on a schedule

Mac Air

6:47 AM daily-brief-cc: calendar, Gmail, Slack, Asana, open commitments, GHL pipeline, Mercury balances, system health. Five parallel Sonnet readers, Fable synthesis, Opus reconciliation audit, gbrain digest.
6:45 AM asana-autorun: runs due [CC-AUTO] Asana tasks in sub-agents, posts evidence, closes verified ones. soak closed 08-31
7:25 AM galarm-watchdog · monthly credits-review.
Cowork sidebar: 4 AM candidate check-in drafts, 6 AM LGH new-member check, Thu 10 AM EOS L10 implementer digest.

Mac mini “CORA”

every 30 min cora-inbox-check · hourly cora-slack-check.
5:00 AM galarm-daily-alarms (needs the mini's Chrome). flaky
nightly gbrain-consolidation-daily.
ho-ops-channel-check moved to Hermes on 2026-08-31.

Hermes VPS

every 10 min gateway watchdog · 5:00 AM health digest to Slack (the brief reads it).
2:00 UTC nightly gbrain dream · nightly cbm-backup to R2 · daily drift monitor.
Mon–Fri 5 AM missing-paperwork report (deterministic Python, DMs Jason and Laura). live since 08-23
every 30 min ho-ops-check (GitHub instruction channel) · every 15 min gbrain relay · egress and OOM guard timers.

Cloudflare Worker crons

every 15 min cora-ops: CTS capture 05/16 UTC, reconcile 10/21 UTC, Europe lead catcher.
2×/day cts-deadman · hourly mercury-balances · 8 AM qb-bookkeeping (CORA Mercury ingest).
every 30 min ghl-email-stats deliverability pager · daily ho-dashboards plus its 15-min watchdog.

The systems with workflows attached

CTS, the Cora Triage System. A self-reconciling commitment ledger in D1. Reads Jason's sent mail, Slack, Granola transcripts, and GHL. CAPTURE upserts promises; RECONCILE closes only on a dated artifact, escalates stragglers to Asana, and DMs an end-of-day summary. Fully cloud-owned since 2026-08-25, 96/96 clean reconcile runs, backlog 126 → 14. Still in candidate mode: precision ≈65% on the July label set, so auto-close authority is not yet earned.
CORA AI Bookkeeper. Mercury and Chase → D1 double-entry ledger → review UI behind Cloudflare Access → Xero as the books. Track 1 (reconciliation integrity) code-complete 2026-08-30, 2,553/2,553 tests, three independent Codex gates before any month locks, $0 penny tolerance. July close awaits operator review.
Daily Brief as above, plus the cloud shadow in soak. Cutover is a deliberate human edit.
Claude Autorun. Asana as a job queue for Claude: [CC-AUTO] runs alone, [CC-ASSIST] needs one human step. Both are filed in a dedicated section hidden from the daily brief.
MCP connectors (Cloudflare Workers, Google-OAuth-gated, allow-listed team): gworkspace-mcp (Gmail, Calendar, Drive, Sheets, Docs, the canonical first path for every Claude surface), ghl-mcp (73 GoHighLevel tools), asana-mcp, sops-mcp (SOP system on D1 + Vectorize), cts-slack-mcp. Web apps: EOS Checkup and Quarterly Reflection, Closed-Deals dashboard. Three websites on Cloudflare Pages auto-deploying from GitHub.
Automation Registry. Week 1 live (Healthchecks, alert rail, drills). Weeks 2–3 deferred; kill review 2026-10-15.
Coordination repos. higher-options-ops (INSTRUCTIONS/ in-box and STATUS/ out-box between hosts that share no filesystem), planning (plans Hermes reads and reviews as GitHub Issues), hermes-ops (DR mirror), and per-product repos for CTS, the bookkeeper, the brief, and each MCP.

How we build

Plan in Fable 5 (Opus 5 fallback) with an interactive eng-manager review. Every plan ships runnable acceptance gates: CHECK and EXPECT, read from the system of record, evidence pasted verbatim (Rule G-11).
Code in Sonnet 5 subagents, worktree-isolated. Kimi K3 codes the same modules in parallel via OpenRouter and a judge picks or grafts. First live dual-coder run 2026-08-30.
Review in Fable, then external adversarial gates: Codex (OpenAI, local CLI, read-only sandbox), GLM-4.6, and Kimi K3 (both through Hermes's second-opinion script, OpenRouter, US-pinned, zero data retention). Verdicts are committed verbatim before any finding is acted on. FAIL loops until PASS; the cutover PR took three rounds, br-closeout is at round seven.
GitHub-first (Rule G-12). PRs are the unit of change, gates become CI, history is the audit trail. iCloud files are mirrors only.
“Built” is not “working” (Rule G-7). Deployed → in verification → closed only after a soak of real unattended runs with an un-ignorable monitor.

Continuous Build Model (CBM). The bottleneck was the human at every phase gate. The CBM harness (built 2026-08-30) runs phases in a loop: fan out coders, run gates, green auto-advances, red spawns a fix agent, stuck parks and wakes Jason, then external review and a PR with receipts. Hard $50/night cap, spend ledger, loop watchdog, dead-man heartbeat, and a stop-line for production, money, outward sends, and destructive changes. Five PRs went through it in four days; migration phases 0–2 are live on Hermes. Standing limit: attended runs only until the controller moves off the sleeping laptop.

Team tooling. Jason: Claude Code (Fable, Sonnet, Haiku by tier) with Codex cross-checks; ChatGPT is used across the team to review Claude output and vice versa. Sandra: websites with Claude Code via GitHub. Jerme: reported by Jason to be coding with Kimi and GLM now (not yet in the ops docs). Grok Bot was evaluated 2026-08-29 and rejected as an ops backbone.

Open questions worth the expert's eye

Where should the overnight build controller live? The laptop sleeps, Hermes has 7.8 GB and no swap, and the cloud sandbox has egress limits.
Earning autonomy. CTS needs a fresh precision eval before auto-close. The same candidate → autonomous ladder will apply to Cora's GHL writes.
Monitoring the monitors. cts-deadman and every mini-hosted routine still lack an absence alert; registry Weeks 2–3 are deferred.
Near-term. gworkspace OAuth bearer expires 2026-09-13 (renewal prepped); daily-brief cloud cutover; finish Mini→Hermes phases 3–5 and power the mini down; br-closeout PR #4 at Codex round 7.