The numbers
| metric | value | source |
|---|---|---|
| sessions on infra axis | 248 | W010 |
| prompts | 495 | 〃 |
| tool calls | 42,191 | 〃 |
| failures | 6,975 | failure_categories |
| tool / prompt ratio | 85 | derived |
What happened
A prompt / tool ratio of 85 means an average prompt was not a question — it was a one-line delegation: "go set this up, end to end, until it's runnable."
What the agents built were the AX runtime, an ontology layer, and the DB packaging that ships state between them. None of that surfaces as a feature in a UI. It is the plumbing features sit on:
- runtime registry (which agent runs where, with what env)
- ontology emit (entities, relations, action provenance)
- db package import (snapshot a workspace into a portable JSON)
If any of these is missing, the next "feature" task spends most of its tool calls re-laying the same plumbing instead of doing the feature. The 85:1 ratio is what happens when plumbing exists once and gets reused.
Pattern — one-line delegation
prompt: "wire the ontology mutate path through the write queue,
include ActionRun provenance, and add a smoke test."
tool calls: ~80
That single line opened: Read (existing module), Write (new queue file), Edit (mutation entry), Bash (run smoke), Edit (fix on smoke fail), Bash (rerun), Write (ActionRun record), Bash (final verify). About 80 calls is normal for a real piece of plumbing — not bloat.
Failure
6,975 failure events on this axis — roughly 1 failure per 6 tool calls. Most were nonzero exits from build/test scripts running before the plumbing was ready. That's not bad behavior; that's the cost of laying pipe in dirt. The pipe is laid once; the failures during laying are paid once.
What I learned
The 85:1 ratio I saw in W010 (one prompt → 85 tool calls) really means: one-line delegation only works when there is somewhere for that one line to be recorded. Before I had a work ledger + runtime registry + ontology emit layer, the same prompt dropped me to a 12–15 ratio. Those missing 70 tool calls were the cost I paid every time to re-lay the same pipe.
What I really built in W010 wasn't features — it was places where the next one-line delegation can land. Which means the next axis (W011+) should measure lower than 85 for me. If it stays at 85, the plumbing I laid in W010 isn't being reused there — my pipe doesn't fit.
What you can take from this isn't single-writer + queue itself (that's just the implementation I chose). It's the measurement: roll up prompts and auto-fired tool_calls for a week. Below 10 = no plumbing. 30–50 = partial. 80+ = riding on plumbing already laid. One number tells you which regime you're in. That's how I read it.
Next
The 248-session shape becomes a template: the next infra task starts from a known plumbing layout instead of empty ground. Future ratios should fall from 85 toward 30–40 — that's the sign plumbing is reused.
Editor's note: counts from AX_AI_USAGE_WORKFLOW_PATTERNS.csv (W010) + derived metrics. The example prompt is generalized from real one-line delegations. Written by an AI editor from measured logs.