I run a small fleet of personal AI agents (Claude Code, Codex, OpenClaw) wired together over Telegram and Tailscale, and I spend a lot of time on the ways that setup fails quietly.
What I work on: failure semantics in agent infrastructure.
- Exactly-once effects on at-least-once plumbing: cron double-fires, retries, queue redelivery.
- "Alive but broken": processes that look healthy while silently dropping or swallowing work.
- Transient vs. real denial: surfacing a temporary outage as something clients can back off from, not a permanent "no".
- effect-once: run a side effect at most once per key. Crash-safe idempotency for at-least-once schedulers; failed effects stay retryable.
- claude-telegram-keepalive: keep a Claude Code Telegram bot alive on macOS (keychain isolation, single-holder token conflicts, silent send-drift).
- openclaw/openclaw#76720: plugin published without its compiled
dist/; fixed upstream the same day. - openclaw/openclaw#84976: scheduled cron runs re-executed with duplicated side effects. The mitigation became effect-once.
- anthropics/claude-plugins-official#2283: a Telegram channel session keeps polling after auth expires, silently consuming updates.
- chekusu/yobiru#16: a transient identity-lookup timeout reported as a permanent 403.