Agents pass work forward through time via scheduled triggers. No two agents coexist.
Status: Primitive (not a standalone mode) | Cost: ~$0.05-0.50/handoff | Complexity: Low
- Work that spans multiple sessions or days
- Tasks that benefit from "sleeping on it" (waiting for data to settle, market close, etc.)
- Overnight batch processing where no human is available
Cross-Time Relay is persistence + scheduling, not a collaboration pattern. GPT-5.4 noted: "LangGraph threads/checkpoints and AutoGen save/load state make this practical — treat it as shared infrastructure, not a flashy standalone mode."
If your platform has cron + checkpoint files, you already have relay capability.
Day 1 15:00 Agent A --> writes findings.md + relay-note.md
Day 1 23:00 Cron triggers Agent B --> reads relay-note.md --> continues work
Day 2 09:00 Operator arrives --> reads synthesized results
The departing agent writes a structured relay note, not free-form:
## Relay Handoff
- **Conclusion so far**: [2-3 sentences]
- **Open questions**: [numbered list]
- **Key artifacts**: [file paths]
- **What NOT to re-investigate**: [already ruled out, with reasons]
- **Suggested next step**: [specific action]The "What NOT to re-investigate" field is crucial — without it, the next agent wastes half its budget re-deriving what the previous agent already explored.
| Scenario | Response |
|---|---|
| Relay state file missing/corrupt | Alert operator, do not proceed blindly |
| Cron skips a trigger (machine sleeping) | Check expected_next_run timestamp; if gap > 2x interval, log and decide |
| Agent produces incomplete handoff | Next agent should flag the gap, not silently continue |