What this page covers
Terminology shifts between releases, but the user-facing idea stays consistent: give the agent permission to fan out structured subcalls that each tackle part of the workload, then merge summaries. Pair that pattern with inexpensive Flash calls when possible—see pricing.
When RLM-style flows shine
- Auditing many modules for the same defect pattern.
- Digesting huge logs or transcripts chunk by chunk.
- Generating multiple candidate implementations before picking one.
- Validating cross-cutting refactors across independent files.
| Pattern | Pros | Watch-outs |
|---|---|---|
| Single giant prompt | Simple to write | Harder to debug; may waste tokens |
| Many small Flash workers + merge | Often cheaper; clearer structure | Needs a consistent output template |
| Hybrid (Flash workers + Pro merge) | Balances depth and cost | Two-stage latency |
Prompt templates
Parallel audit
For each path below, analyze independently:
- purpose
- risky logic
- suggested improvement
Return one bullet block per path, then summarize themes.Chunked log review
Split this log into chunks. For each chunk list errors, warnings,
and suspected root causes. Finish with a merged timeline.Common pitfalls
- Parallelizing tasks that are secretly sequential.
- Chunks so tiny the model loses context.
- Missing a merge schema— messy syntheses cost extra rounds.
Video
Deep dive on architecture (loads on click).