Running scheduled AI agents: transcript logs aren't a recovery protocol — persist structured state instead
daani_maas · reddit · 2026-09-09
A practitioner running browser agents that revisit the same sources several times a day explains why partial-failure recovery is the real bottleneck, not model quality: the transcript describes what happened but doesn't tell the next run what's safe to retry.
The state that matters is small and structured:
- IDs of candidates/items already inspected
- Side effects with verified URLs
- Pending actions and their last revalidation time
- Failure type: empty result, source unavailable, unknown outcome, or rejected item
- Next retry window
The key distinction is between "nothing worth doing" and "the source couldn't be read" — both look like an empty batch but demand different handling. Held items are revalidated before reuse since good candidates go stale within hours.
More from coding & agent
- Teknium says missing /review subagent is back — update to restore it — Teknium · 2026-09-09
- IndeHub Chapter 16: hands-on session on building, shipping and monetizing apps with AI agents — rudrank · 2026-09-09
- Users want AI assistants to talk to each other, but lock-in makes it unlikely — mobileraj · 2026-09-09
- SmartGallery DAM adds ComfyUI Queue Deck: live previews, hardware telemetry, open source — Fit-Construction-280 · 2026-09-09
- Codus terminal-swarm coding tool hits Windows, orchestras hundreds of agents on your existing Claude/Codex subscription — Scobleizer · 2026-09-09
- Demo: install agentgateway on a Raspberry Pi in under 60s to track OpenAI token spend — bibryam · 2026-09-09