Token-rich vs Token-poor Loops in Loop Engineering
Short answer
Token-rich loops pass large context into every iteration. Token-poor loops pass summaries, retrieved memory, or scoped observations to reduce cost.
Why it matters
Context is multiplied by every iteration and every run. A loop that re-sends a huge prompt each turn is easy to build but expensive and more likely to hit context limits; a token-poor loop is cheaper and faster but needs real memory and state management.
Practical checklist
- Estimate input tokens per iteration before scheduling
- Decide what to keep, summarize, retrieve, and forget
- Pass only the observations relevant to the next action
- Re-estimate cost when iterations or runs/day increase
Example
Instead of re-sending an entire repository each iteration, a token-poor loop passes a summary of the last error and the few files involved — cutting input tokens from 80k to 8k per turn.
Common failure modes
Re-sending full context every iteration
No summary or retrieval strategy
Cost grows silently as runs/day increase