liveBook · Manning

Re-read ratio

As explained in Designing AI Agents

A metric measuring how often an agent retrieves the same file or information more than once during a task.

The re-read ratio tracks how often an agent retrieves a file or result it has already seen. It is one signal of perception quality: repeated reads can indicate that useful information was deferred or summarized too aggressively. section 3.1.4

Use the metric to investigate retrieval behavior before changing context budgets:

  • Establish a baseline by task type and retrieval policy. section 3.1.4
  • Investigate sustained increases, especially after triage or compaction changes, before tightening the budget further. section 3.1.4
  • Derive the metric from tool-call traces; change one perception policy at a time and check task success, latency, and cost together. section 3.1.4

How Designing AI Agents covers it

How deep it goes
Track reads against task-type and retrieval-policy baselines; investigate sustained increases before tightening context budgets.
Examples
Compare repeated reads of files or results, especially after triage or compaction changes.
What it assumes
Derive counts from tool-call traces and establish workload baselines by task type and retrieval policy.
What you can do after
Diagnose possible over-deferral or over-compression by inspecting repeated reads before reducing retrieval or context budgets.
How it connects
Connects to perception quality, semantic compaction, triage, token spend per successful outcome, and attention-zone coverage.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.