I asked my AI agent to find problems while I sleep. It works.

Every night at 3:00 AM, my Hermes agent reads the entire Obsidian vault — every project file, goal document, SOP, trading note, and system log — generates ideas for things that need building, picks the best one, builds it, tests it, and reports back.

I wake up to a working system I didn't ask for.

Here's how it works and what it's actually built.

---

The setup

The "nightly-obsidian-ideas" cron job runs daily at 3:00 AM SAST. The prompt is simple:

1. Read all files in the Obsidian vault (_context/, trading/, reference/)
2. Identify gaps, missing tools, or improvement opportunities
3. Generate 3 ideas with brief descriptions
4. Select the highest-impact idea
5. Build it — write the code, create the cron job, write the documentation
6. Test it
7. Report to Discord with what was built and why

The agent has full access to the terminal, file system, and cron scheduler. It can create scripts, set up scheduled jobs, and modify configuration files.

---

What it's actually built

Website uptime monitor (Night 1). The agent noticed that while we were monitoring Wix form emails, we weren't monitoring whether the websites themselves were online. It built wix_uptime_monitor.py — a script that pings comfortshooting.com and thatslekker.co.za every 15 minutes and sends Discord alerts on downtime. It created the cron job, wrote the Obsidian documentation, and reported both sites as UP (710ms and 378ms response times).

XAUUSD trade journal (Night 3). The agent identified that we had automated trading briefings but no way to log actual trades or track performance. It built xauusd_trade_journal.py — a trade logger with add/summary modes, JSON storage, win rate calculation, average R&R, P&L tracking, streak counting, and max drawdown. Created the cron job (daily 19:15 prompt), wrote the documentation, and tested empty state and trade entry.

XAUUSD session prep pinger (Night 4). The agent noticed that the morning briefing covers the previous day's action but there was no pre-session checklist. It built xauusd_session_prep.py — a script that runs at 17:45 (15 minutes before the US session opens), fetches live gold price from 3 sources, loads S/R levels, checks the economic calendar, and outputs a pre-session checklist. Tested successfully with a live price of $4,541.20.

XAUUSD risk calculator (Night 5). The agent found that while we had risk rules documented, there was no tool enforcing them pre-trade. It built xauusd_risk_calculator.py — position sizing from account balance + risk percentage + stop-loss distance, with enforcement of the 3% daily loss limit, max 2 trades per session, and minimum 1.5:1 R&R. Reads S/R levels from the morning briefing's output. Daily state auto-resets at midnight.

---

What it tried to build but didn't finish

SEO sitemap pipeline (Night 6). The agent identified That's Lekker needed automated SEO monitoring and started building a sitemap pipeline. The build was truncated mid-execution — the session ran out of time before the script was complete. This is logged as "in progress" in the vault.

---

How the idea generation actually works

The agent isn't guessing. It reads the actual project files:

• goals.md — which tracks progress across 5 goals with checkboxes
• comfort-shooting-labourless-goal.md — the detailed 4-phase automation plan
• hermes_activity.md — the day-by-day log of what's been built and what broke
• xauusd-strategy.md — the trading strategy documentation
• Various SOP and reference files

When it sees a checkbox that's unchecked, a documented gap, or a manual process that could be automated, it generates an idea to address it. The ideas are specific and grounded in actual project state — not generic suggestions.

---

The failure mode

The system isn't perfect. On one run, the prompt had been changed (by me) to only generate ideas without building anything. The agent dutifully generated 3 good ideas and stopped. I noticed the next morning, updated the prompt to include the build step, and it's been building ever since.

The lesson: the agent does exactly what you tell it to. If the prompt doesn't say "build it," it won't build it. Prompt design matters as much as the automation itself.

---

Why this matters

Most AI agent demos show the agent responding to a prompt. This is different. This is the agent proactively identifying problems in my actual business, building solutions, and deploying them — without being asked.

I come to new tools every morning. Tools I didn't design, didn't spec, didn't request. Built from reading my own project files and finding the gaps.

That's the difference between using AI as a tool and deploying AI as an operator.

---

The nightly-obsidian-ideas cron job runs daily at 3:00 AM SAST. All tools described here are running on the live system. The Obsidian vault is at the source of every idea — nothing is generated from scratch without checking actual project state first.