liveBook · Manning

Token count

As explained in Vibe Engineering

Token consumption is the quantity of language-model input and output tokens used by an AI system or user over a defined period.

Token count measures the input and output tokens consumed by an AI system or user over a defined period. Treating it as a productivity achievement creates tokenmaxxing: teams burn more tokens because the metric rewards volume, not shipped value. section 9.5.6

The disciplined position is neither “minimize tokens” nor “maximize tokens.” Lower per-interaction cost can expand usage into reviews, documentation, commit messages, tests, and experiments. In one case, interactions rose 54% while the monthly bill still fell 84%. section 9.5.6

Use token count as an operational signal, not a success measure:

  • Compare consumption with shipped outcomes; token volume alone cannot establish productivity. section 9.5.6
  • Monitor anomalies before invoices arrive, especially across parallel agent conversations. section 5.6.2
  • Trace spikes across agent boundaries because retries may burn tokens without correcting systematic failures. section 5.6.2

Agent pipeline dashboard

A payment-retry pipeline reports 47k tokens and $0.42 consumed, with 35k tokens estimated to remain. The same view exposes stalled implementers, waiting testers, and queue bottlenecks, turning token count into evidence for investigation rather than a leaderboard score. section 5.6.2

Pipeline: payment-retry-feature
Status: Running (7m 23s)
[✓] Architect      (2m 14s, 12k tokens)  → design.md
[✓] Implementer-1  (3m 02s, 28k tokens)  → resilience.py
[◐] Implementer-2  (4m 11s, running)     → gateway changes
[ ] Tester-1       (waiting)
[ ] Tester-2       (waiting)  
[ ] Documenter     (waiting)
 
Total tokens: 47k ($0.42)
Estimated remaining: 35k tokens, 5 minutes

Inspect token consumption alongside agent status, elapsed time, cost, artifacts, and estimated remaining work. section 5.6.2

How Vibe Engineering covers it

How deep it goes
Explains metric misuse, rebound effects, cost arithmetic, and dashboard monitoring without descending into model internals.
Examples
Uses token leaderboards, cost comparisons, interaction growth, and an agent-pipeline dashboard as concrete examples.
What it assumes
Assumes familiarity with monitoring services and infrastructure before extending observability to agents.
What you can do after
Diagnose/debug with it by correlating token spikes, agent failures, queue delays, costs, and remaining work.
How it connects
Connects token count to productivity metrics, the rebound effect, observability, and agent pipelines.

Also explained in

Compare how 6 books explain token count →

Related concepts

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