Eve, Codex and Claude Code: scheduled agent setup

The speaker explains his setup for recurring business automation at Inbox Zero. Each AI agent works toward a measurable objective, checks results after a suitable interval, and records what it tried. His examples include reviewing SEO changes after two weeks and checking advertising results after three days. He recommends testing a task manually before scheduling it.

The setup discussion covers desktop schedules, GitHub Actions and Eve. The speaker says desktop jobs depend on the computer being available and describes previously running Codex through GitHub Actions with the required credentials. He later moved to a self-hosted Eve instance on Hetzner so he could continue conversations with agents through Slack rather than receive webhook updates alone.

His agents share one Git repository, with a core agent delegating to specialist subagents. They commit their findings so later runs, and work through a coding assistant on his computer, can use the same history. A memory.md file records experiments and their outcomes.

The final sections cover browser automation and restricted access to business data. The speaker recommends read-only permissions and limiting exposed columns. He also describes subscription authentication workarounds for reducing separate API charges, while acknowledging experimental support and possible blocking. These are reports about his setup, not guarantees of ongoing availability or cost.