SERIES RECAP · ECONOMICS OF CLAUDE CODE
The helper scorecard, two months later
Nine helpers shipped. Three absorbed into the harness. One new. The math on the rest still holds.
| # | Helper | Status | Verdict |
|---|---|---|---|
| 01 | Idle Tax | Math Stands | 12.5× re-cache is real; the prescription softened, and idle-autosave now closes its loop. |
| 02 | Just One More Turn | Keep | Context rot is real; split at natural boundaries. The knee at ~50 turns still holds. |
| 03 | Subagent Isolation | Absorbed | Delegation guidance is now harness-native; the 50-file nudge stays as a heuristic. |
| 04 | Compact Gamble | Reframed | Compaction now summarizes structurally rather than truncating; the marker stays as audit trail. |
| 05 | Watching Cost | Absorbed | The harness auto-persists oversized output natively; the mental model stays useful. |
| 06 | Delegation Cost | Keep | Still the only sensor for agent-result bloat. No harness equivalent. |
| 07 | Effort Control | Keep | Deliberate spend is still deliberate. Model selection and effort flags remain manual. |
| 08 | Auto-Persist | Keep | The continuity core. Zero tokens; survives any session boundary or crash. |
| 09 | Idle Autosave | New | Handoffs write themselves while you're away. Closes the loop on the Idle Tax (~$0.01). |
Keep math and prescription both hold
Math Stands numbers correct; prescription evolved
Reframed updated framing, same underlying signal
Absorbed behavior now in the harness; helper retired
New added in this recap
Still active
5
Keep + Math Stands + New
Absorbed or reframed
3
Harness now handles the mechanical parts
Reframed
1
Compact Gamble — better model, same warning