Skip to content

The shift

Five things that changed when the tooling changed, and the number behind each one.

One project, one operator. Everything below is drawn from a single body of work over 117 days. It is a description of what happened here, not a claim about what happens generally.

From teams to soloists

10 roles worth of surface area, covered by 1. The ratings below are keyword coverage over the session transcripts, so they show breadth of work touched rather than depth of expertise.

lowerhigher

Domain coverage0 to 100
Coverage by domain
AI / ML79
Backend75
Security65
Systems59
Frontend52
IoT / Edge48
Data Engineering47
DevOps40

From sprints to streams

Throughput117 days
Commits29,024
Repositories touched135
Issues opened4,649
Issues closed3,130

Issues closed trails issues opened by 1,519. Throughput went up; the backlog still grew.

From meetings to messages

378,808 messages across 129 sessions and 71 projects. The unit of coordination stopped being an hour on a calendar.

The cache effect

98.4% of input came from cache. That is the whole reason a long-running context is affordable: re-reading the project each turn is what the cache is paying for.

Contextcumulative tokens
Cache read88.5B
Fresh input4.0M
Output266.8M

Compound velocity

Tools built for one project get used by the next. That is the part that compounds, and it is also the part this page cannot measure: there is no counter for a script that saved an afternoon six months later.

The cost side of the same window is on cost analysis, including what is measured there and what is modelled.

cost-model.json + ai-pilot-data.json, generated Sep 9, 2026