Operations record Jake O'Donnell, P.E. Charleston, SC Day 137

On February 25, 2026, I sat down with an AI agent for the first time.

I am a mechanical engineer who sells carbon capture projects. This page is what happened in the 137 days since, and the numbers on it are republished from the metrics log of the automation fleet they describe. Nothing here is a mockup.

01 · The fleet, live

fleet-data.json generated Jul 11, 6:45 AM ET source: cron-metrics.jsonl last 24h: 123 runs, 120 ok
Runs logged
10,599
2026-03-19 to 2026-07-11
Active jobs
63
122 names all-time, experiments retire
Success rate
89.6%
every run counted, no edits
Days since day one
137
first session 2026-02-25
Scheduled runs per day, trailing 70 days
0125250jun 24: regression traced + revertedJunJul

The rust you can see in May is a backend migration silently failing. The metrics log caught it: May ran at 23.7% failures, June at 2%, and the June 24 revert closed out the root cause for good. I left the bad weeks on the chart because a fleet you can trust is one whose owner does not hide its incidents.

View as table
Week ofRunsCompletedFailed
2026-05-031051488563
2026-05-10883656227
2026-05-1777673640
2026-05-2466162734
2026-05-3173571817
2026-06-0773071812
2026-06-1473572015
2026-06-2176174714
2026-06-2878976623
2026-07-0573471222
What the fleet actually does · 3,125 substantive runs by category
LinkedIn ops
1,653
Data & file pipelines
623
Autonomous build loops
349
Blogs & web content
178
Briefings & news
96
Personal ops
79
Email & inbox ops
67
Health, backups & sys
39
Ad-hoc tasks
30
KPI & reporting
8
Other
3

+ 7,474 lightweight scheduler ticks (a 15-minute content-queue check) excluded above so the real work is readable.

02 · The log

2026-02-25
First session ever with an AI agent. A self-hosted framework running a chief-of-staff agent over Telegram. Day-one output: a complete LinkedIn launch strategy for my company and an overnight research assignment. I did not sleep much that week.
2026-03-12
Company website live in production. First commit and launch, day 16. Still running, still publishing.
2026-03-19
First scheduled job fires. Honestly, it was a golf tee-time watcher. The serious ones followed within the week.
2026-03-24
Cut everything over to scheduled, headless Claude Code on macOS launchd. Outgrew the agent framework in 27 days. This decision is the backbone of the whole fleet.
2026-03-27
Estimating tool scaffolded. 185 historical estimates extracted from OneDrive the next day. It grew to 608 indexed estimates with comparables and a line-item cost database.
2026-04-13
Golf app: 712 commits in three weeks. Live-money settlement and real tee-time booking, with a kill switch, spending ceilings, and cancel-window safety. Automation that spends money teaches you about guardrails fast.
2026-04-24
A backend migration silently degraded the fleet. The failure numbers in the chart above are real and I left them there on purpose.
2026-05-18
Marketing portal first commit. Next.js 15, Postgres, SSO, end-to-end tests. A gated team portal my colleagues use every week.
2026-05-31
Life Caddie ships. A personal iOS app for golf, meals, and briefings. Its backend runs on my desk behind Tailscale, managed by the same launchd fleet.
2026-06-24
Traced and reverted the regression. Fleet failure rate: 23.7 percent in May, 2.0 percent in June. The instrumentation caught what my attention missed, which is exactly why the instrumentation exists.
2026-07-01
Fleet self-audit. 9,453 runs analyzed, reliability findings ranked and filed. The fleet now reviews itself on a schedule.
2026-07-11
This page ships. The fleet republishes its own numbers here. If you are reading a stale copy, that is a bug report I would genuinely like to receive.

03 · Systems in production

Automation fleet

Scheduled headless Claude Code jobs: LinkedIn drafting and replies for four brands, blog pipelines, executive briefings, KPI reporting from plant HMI screenshots, health checks on all of it.
68 active jobs · 10,000+ logged runs
launchd · Claude Code · Playwright

Estimating tool

Self-serve estimating app for an industrial equipment builder. Historical estimates extracted from OneDrive into comparables and a line-item cost database.
608 estimates indexed, 2015 to 2026
Python · Node · Vercel

BD pipeline platform

Real-time deal dashboard with live financial models and proposal, deck, and proforma generation. The sales side of the house runs on it.
Live in production since March
Alpine.js · Vercel · Redis

Marketing platforms

Four production websites with automated content pipelines, plus a gated team portal with SSO, review workflow, and CRM lead capture.
4 sites · 300+ combined commits
Next.js 15 · Postgres · HubSpot

Life Caddie iOS

Personal life-ops app for golf, kitchen, and briefings with push notifications. Backend on a Mac behind Tailscale, supervised by the fleet.
3 launchd services: api, health, backup
SwiftUI · Node · APNs

Autonomous build loops

Scheduled jobs whose only duty is improving the other tools: overnight buildout and polish passes that ship reviewed commits by morning.
340+ build-loop runs logged
launchd · Claude Code · git

04 · The other half

Before any of this I was the opposite of a software person. Licensed mechanical P.E. in North Carolina and South Carolina, HVAC design on pharmaceutical cleanrooms, then capital equipment sales, roughly $3M in revenue in 18 months. Today I am VP of business development at a carbon capture company: I took an industrial CO2 plant from construction to beverage-grade qualified production, and I carry the offtake LOI book. The fleet exists because I needed leverage, not because I needed a hobby.

Small teams are home. In my day job I am the BD lead, the plant-startup engineer, and the toolsmith, usually in the same afternoon.

What I bring is the pairing: I can build the AI employee, and I can also sit across from the client who is buying it, scope it in their language, and close. I know what a cabinetry shop's estimator actually worries about, because I have been an estimator and I have sold to them.

I want to help build Charleston AI.

Building and delivering client AIs, teaching the classes, closing the accounts, or all three. I live in Point Hope, a few minutes from Jessen Lane, and the fleet is live any day you want to see it run.

781.534.0355 · linkedin.com/in/jakeodonnell1