01What we actually built
Apex was a tool factory. Its charter was simple: build software Parallax could use or sell. Security audit tools, site analyzers, content generators, client-reporting workflows. On a fixed 8-hour cycle it read its manifest and its own evolution history, asked one AI model to plan the next build, dispatched the work through a multi-agent orchestrator we called Forge, and had a different model cross-review the result before auto-merging and deploying it.
The rule that made it an experiment instead of a demo: humans never wrote code in its repository. We set the goal, built the harness, and watched.
02The run
First manifest committed. Cycle 1 plans and ships its first tool.
A git-hosting quirk plus an HTTP redirect silently dropping POST bodies broke pull-request creation for days. Apex kept planning; nothing merged. The fix taught us its "definition of done" was too loose, so we rewrote it: every tool must ship as a single self-contained HTML file that works from a USB stick.
Cycle 312. Apex reconciles its own inventory: 134 tools, 74 workflows. The catalog includes an invoice-generator configurator, DMARC and deliverability checkers, and a security-audit suite.
Cron disabled. The experiment answered its question and attention moved to shipping products.
A host reboot kept resurrecting one replica; Apex quietly reached cycle 415 before failing its own health check. We scaled it to zero and pinned that in the stack config. Dormant, not deleted.
03What it got right
The build-review-merge loop between two different AI models ran unattended for months and shipped genuinely usable artifacts. The proof wasn't a benchmark. It was a USB stick: eight self-contained security tools that got carried into real client meetings and opened straight from the drive, no hosting, no login, no excuses.
04What it got wrong
- It hoarded work. Apex opened a tracker task every cycle and closed none of them. A human eventually cleaned up 28 stale tasks. Autonomous systems need janitorial instincts, and they don't develop them on their own.
- "Off" wasn't off. Disabling its schedule didn't survive a host reboot; a replica kept self-restarting. If you run autonomous systems, the shutdown path deserves the same engineering as the build path.
- Its self-reported numbers drifted. The inventory it announced at cycle 312 never fully reconciled with its own manifest. That's why the figures above are labeled verified or self-reported. We hold our tools to the same standard we'd hold a vendor's.
05What it became
The factory is dormant. Its products clocked in this morning. Apex-built tools run in production daily behind our consulting work: the free email security check on our main site came off its line, and an internal service still loads its whole catalog, every tool one API call away, for audits, mail-posture checks, and client reporting. The experiment ended; the inventory went to work.
If you want the same discipline pointed at your business, that's the day job: Parallax Intelligence.