Postmortem: an autonomous outreach agent double-sent emails
eazyigz123 · reddit · 2026-07-21
Postmortem: an autonomous outreach agent double-sent emails
The author describes a failure in a Mac-based autonomous outreach loop that sent duplicate emails at midnight, along with other template errors such as lowercase first names and leaked internal CRM slugs.
What went wrong
- Two agent sessions drained the same pending-send queue.
- The queue had no claim/lease semantics, so both sessions treated the same items as ready.
- This is framed as a classic distributed-systems bug, except the blast radius is real human trust.
What changed
- Auto-send was disabled until sends go through a dedupe ledger checked at send time, not queue time.
- Human-review flags are enforced at the send boundary, not just during planning.
- Quality gates now fail closed: if a template renders a lowercase name or internal slug, it should not send.
Open question
The author asks where people gate side effects in real-world agents: at the queue, the executor, or a proxy in front of the action itself—and how they dedupe when multiple machines or sessions can run the same work.
More from coding & agent
- Greptile routes code review to the other model when it detects GPT or Claude-generated PRs — ycombinator · 2026-07-21
- AI can pass every test and still fail in production, the post argues — JnBrymn · 2026-07-21
- Anthropic says 65% of product-engineering PRs now close through Slack-based Claude Tag — alliekmiller · 2026-07-21
- Google DeepMind rolls out Gemini 3.6 Flash, 3.5 Flash-Lite and Flash Cyber — GoogleDeepMind · 2026-07-21
- AI agents are starting to strain code hosting platforms — craigsdennis · 2026-07-21
- Omnigent 0.6.0 adds Claude Code imports, Slack approvals and desktop apps — matei_zaharia · 2026-07-21