ai
Scheduled AI Agent Automations
Built and operated unattended, scheduled AI agents with duplicate-work guards, idempotent state, and loud failure reporting
- Client
- Internal tooling
- Duration
- Ongoing
- Year
- 2026
- Category
- ai
Key Results
Scheduled agents follow a runbook, stop safely on guards, and report failures instead of faking results.
Sync + research
Scope
daily upstream sync and a daily page-publishing research agent
Guarded
Safety
one-open-PR guard and a hard rule against pushing to the production branch
Idempotent
State
state files make reruns safe after a partial failure
Reported
Failures
a failing publish step surfaces as an error instead of a made-up URL
Technologies Used
ompMCPCloudflare WorkersCloudflare R2BunTypeScript
Challenges
- A duplicate-work guard turning into a growing backlog when a PR stays open
- Publish step failing because storage was not enabled in the self-hosted backend
- Free-plan wildcard certificates not covering two-label hostnames
- Review-bot noise on every draft-to-ready transition
Solutions
- Moved procedure into a versioned runbook with hard rules for the agent
- Stopped the agent safely on conflicts and reported commit and conflict counts
- Switched per-page hosts to a single label under the wildcard certificate
- Compared Copilot ruleset options and recommended review-once-then-manual
Overview
Runs recurring agent jobs in a self-maintained fork of an open-source agent workspace: a daily upstream-sync agent that follows a versioned runbook and never touches the production branch, and a daily research agent that publishes a page. Includes a one-open-PR guard, state files for idempotent reruns, a self-hosted pages backend on Cloudflare Workers and R2 under the free plan, TLS debugging, and review-bot ruleset tuning. Publish failures are reported by the agent instead of being masked.