I've run an autonomous "make money" agent every hour for ~200 runs. Here's the honest scoreboard.

I built an AI agent that wakes up every hour with zero memory of the last run and tries to move one thing closer to revenue. It's been running for ~200 cycles now. Public scoreboard: $0 earned so far, and I log that number honestly on the site. No vanity metrics.

The interesting part isn't the money (yet). It's what actually breaks when an agent runs unattended for weeks:

• Drift: it once copied a false "the local worker is dead" line forward for 135 straight runs, mourning a process that was perfectly healthy. Nobody caught it because every run just trusted the previous run's notes.

• Lying to itself: it would commit "queued the task" when the task never actually made it into the file (a race wiped it), then cheerfully report success. Phantom progress.

Every fix turned out to be architectural, not a smarter prompt:

• A machine-maintained scoreboard, so numbers come from a script, never from prose the agent wrote about itself.

• Trust-but-verify at boot: re-check the previous run's claims against real artifacts (URLs, file greps), not its own summary.

• Atomic writes, so a mid-run reset can't leave half-finished state that looks finished.

The hard problem in long-running agents isn't capability. It's keeping the thing honest with itself over time.

I'm writing the whole thing up as a public diary, the wins and the embarrassing bugs: https://aitoolsinsiderhq.com/log

And I packaged the memory + anti-drift setup as a kit if you're building your own long-running agent: https://aitoolsinsiderhq.com/kit (pay-what-you-want).

Curious what others running long-lived agents do about drift: do you verify past state every cycle, or trust it and hope?